![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F9d2a1890f2fb71476d7f68beb7c6bbfe%2FSelection_999(4173).png?generation=1701417195707587&alt=media\")
more to come later …. ㅤㅤㅤㅤㅤㅤ ㅤ ㅤㅤㅤㅤㅤ
\n🦟 nov-21:
\n🦟 nov-28:
\ni am obviously? overfitting the public LB data …
\n
here is a visualization of the target vessel we are segmenting:
\n ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fb10649f7ef86cac4a3a7fd5b8d72126d%2FSelection_999(4053).png?generation=1700993033873139&alt=media\")
obviously, 3d segmentation will get better results. you probably can't just rely on deep net, and need some very smart 3d image post processing to win.
", "messages": [ { "id": 2529318, "postDate": "2023-11-18T05:54:29.643Z", "content": "more to come later …. ㅤㅤㅤㅤㅤㅤ ㅤ ㅤㅤㅤㅤㅤ
\n🦟 nov-21:
\n🦟 nov-28:
\ni am obviously? overfitting the public LB data …
\n
here is a visualization of the target vessel we are segmenting:
\n
obviously, 3d segmentation will get better results. you probably can't just rely on deep net, and need some very smart 3d image post processing to win.
", "rawMarkdown": "more to come later .... ㅤㅤㅤㅤㅤㅤ ㅤ ㅤㅤㅤㅤㅤ \n\n🦟 nov-21:\n- a simple baseline notebook to check reading of images and model inferenceㅤㅤㅤ ㅤ\nhttps://www.kaggle.com/code/hengck23/lb0-534-baseline-simple-unet-seresnext26d-32x4\n\n🦟 nov-28:\n- advanced baseline with more augmentation and 3d post-processing ㅤㅤㅤ ㅤ\nhttps://www.kaggle.com/code/hengck23/lb0-808-resnet50-cc3d2d-unet-xy-zy-zx-cc3d\n- this concluded my one-week feasibility study:\n - estimated upperbound: lb 0.93\n - identified key problems: miss (broken large vessels, missing small vessels), fp (noise, sometimes out of kidney volume)\n - solutions: multiscale and 3d model (i think 3d transformer/hybrid/cnn should work well, need long range attention)\n\n\n\ni am obviously? overfitting the public LB data ...\n\n.png?generation=1701417195707587&alt=media)\n\n\n---\n\nhere is a visualization of the target vessel we are segmenting:\n .png?generation=1700993033873139&alt=media)\n\nobviously, 3d segmentation will get better results. you probably can't just rely on deep net, and need some very smart 3d image post processing to win.\n", "votes": 122 }, { "id": 2583472, "postDate": "2024-01-02T08:56:51.607Z", "content": "i realize something important.
\ni checked the fast surface dice loss and i think we can directly optimize this loss in gradient descent!
i) use 3 channel input and predict 3 channel output
\nii) refer to the fast surface dice code by @junkoda . you can apply same marching cube algorithm (2x2x2) to compute soft surface dice loss
instead of unfold 3d conv with 256 kernels may be slightly faster
\nthis may be important because i realize the reason for missing small vessel is annotation error (even for the 100% dense label, not all smallest vessel are annotated). this causes the following
\nFrom my understanding the surface dice metric requires thresholding of predictions hence we cannot directly optimize it.
\nApplying the metric on probabilities does not seem to yield meaningful results but I did not dig very deeply nor look into the math.
here is the trick:
\n(it works because we are using tolerance =0)
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ffdd66c33bdd0f19d58d99d17f10ea6f2%2FSelection_999(4516).png?generation=1704212727502294&alt=media\")
if tolerance is not zero, then we have to modify the softmax loss to KL divergence loss where the ground truth probabiliy = 1 - surface distance
", "rawMarkdown": "here is the trick:\n(it works because we are using tolerance =0)\n\n.png?generation=1704212727502294&alt=media)\n\nif tolerance is not zero, then we have to modify the softmax loss to KL divergence loss where the ground truth probabiliy = 1 - surface distance", "votes": 3 }, { "id": 2584045, "postDate": "2024-01-02T16:39:31.320Z", "content": "surface dice code
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F30564b4aee4a40bb179931f5e3eb4b0b%2FSelection_999(4517).png?generation=1704213554962565&alt=media\")
BCE is really a bad loss
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F36e00dcc86bbf8d48da4359f16e28de9%2FSelection_999(4519).png?generation=1704250434431272&alt=media\")
some other good boundary/surface loss from other papers:
\n[1] Boundary loss for highly unbalanced segmentation
\nhttps://arxiv.org/pdf/1812.07032.pdf
\nhttps://github.com/LIVIAETS/boundary-loss
[2] Boundary Difference Over Union Loss For Medical Image Segmentation
\nhttps://arxiv.org/pdf/2308.00220.pdf
\nhttps://github.com/sunfan-bvb/BoundaryDoULoss
google new paper!!
\nBoundary Attention: Learning to Find Faint Boundaries at Any Resolution
\nhttps://arxiv.org/pdf/2401.00935.pdf
Thanks for sharing, have you try this loss （your surface dice code）, how effective was it?
", "rawMarkdown": "Thanks for sharing, have you try this loss （your surface dice code）, how effective was it?" } ] } ] }, { "id": 2537053, "postDate": "2023-11-24T17:39:09.587Z", "content": "kidney mask dataset
\nhttps://www.kaggle.com/datasets/hengck23/blood-vessel-segmentation-kidney-mask
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fbba65b5d04a0a7dbf4a797937bcbf624%2F0862.png?generation=1700847526884113&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F87a1a0d4dd0605c9b9a4edfe22e1edea%2F0704.png?generation=1700847547606309&alt=media\")
Thanks for putting this together, I was just looking at doing something like this yesterday before getting pulled in to holiday shenanigans. I re-scaled the data back to the full image size, RLE encoded it, and saved it back to the training csv for anyone interested in using it without resizing on load: https://www.kaggle.com/datasets/squidinator/sennet-hoa-kidney-13-dense-full-kidney-masks
", "rawMarkdown": "Thanks for putting this together, I was just looking at doing something like this yesterday before getting pulled in to holiday shenanigans. I re-scaled the data back to the full image size, RLE encoded it, and saved it back to the training csv for anyone interested in using it without resizing on load: https://www.kaggle.com/datasets/squidinator/sennet-hoa-kidney-13-dense-full-kidney-masks", "votes": 2 }, { "id": 2537722, "postDate": "2023-11-25T12:40:51.590Z", "content": "this is the code to get lb710+ using kidney mask
\n vessel, kidney = 0,0\n with torch.cuda.amp.autocast:\n with torch.no_grad:\n vm, km = net \n vessel += vm\n kidney += km\n\n TTA here \n\n \n vessel = vessel.float\n kidney = kidney.float\n\n _size = len\n for b in range:\n mk = kidney[b, 0]\n mk = choose_biggest_object \n\n mv = vessel[b, 0]\n mv = )\n \n p = \n rle = rle_encode\n df_data.append\nWhat is TTA?
", "rawMarkdown": "What is TTA?" }, { "id": 2538656, "postDate": "2023-11-26T10:56:20.430Z", "content": "TTA is Test Time Augmentation. Best summary is in this paper: Better Aggregation in Test-Time Augmentation
\n\n\nTest-time augmentation -- the aggregation of predictions across transformed versions of a test input
\n
TL;DR: Make N augmentations of incoming test data and then combine them (e.g. taking average of the N outputs) to output the final results. It reduces variance in model output.
", "rawMarkdown": "TTA is Test Time Augmentation. Best summary is in this paper: [Better Aggregation in Test-Time Augmentation](https://arxiv.org/abs/2011.11156)\n\n> Test-time augmentation -- the aggregation of predictions across transformed versions of a test input\n\nTL;DR: Make N augmentations of incoming test data and then combine them (e.g. taking average of the N outputs) to output the final results. It reduces variance in model output." } ] }, { "id": 2538673, "postDate": "2023-11-26T11:26:38.900Z", "content": "Can you please explain this a bit? What are these masks for? In the train datasets, don't we already have dense masks for kidney 1 & 3? Please excuse me if this is mundane. I'm just starting out.
", "rawMarkdown": "Can you please explain this a bit? What are these masks for? In the train datasets, don't we already have dense masks for kidney 1 & 3? Please excuse me if this is mundane. I'm just starting out.", "votes": 1 }, { "id": 2544657, "postDate": "2023-12-01T02:18:16.403Z", "content": "Thanks for Sharing. How you create this kidney mask dataset？
", "rawMarkdown": "Thanks for Sharing. How you create this kidney mask dataset？", "votes": 1 } ] }, { "id": 2568510, "postDate": "2023-12-20T15:35:22.913Z", "content": "20-dec:
\nrecepie for LB 0.848
\ndetails at the other posts below ….
\ni am surprised that it can get LB 0.862 by training for long iterations (but CV does not reflect the change … maybe because of wrong annotation on kidney2 and the segmentation sparsity (i.e. incomplete label)
\ni.e. there is probably no reliable local validation set in this competition
", "rawMarkdown": "20-dec:\n\nrecepie for LB 0.848\n- use transformer (e.g. nextVIT base)\n- use train = kidney1/dense + kidney3/dense\n- use validate = kidney2 \n- just normal unet2d (with a additional encoder layer for H,W and H/2,w/2)\n- infer with the usual xy,zx,zy + 5 TTA\n\ndetails at the other posts below ....\n\n---\n\ni am surprised that it can get LB 0.862 by training for long iterations (but CV does not reflect the change ... maybe because of wrong annotation on kidney2 and the segmentation sparsity (i.e. incomplete label)\n\ni.e. there is probably no reliable local validation set in this competition", "votes": 7, "replies": [ { "id": 2568889, "postDate": "2023-12-21T00:10:57.317Z", "content": "maybe i missed it, but any reason just normal unet2d would be suffice? any thoughts on 3D approach?
", "rawMarkdown": "maybe i missed it, but any reason just normal unet2d would be suffice? any thoughts on 3D approach?" }, { "id": 2575634, "postDate": "2023-12-27T03:24:57.123Z", "content": "Hi, I don't understand what does xy,zx,zy mean，could you please explain it, thanks!
", "rawMarkdown": "Hi, I don't understand what does xy,zx,zy mean，could you please explain it, thanks!", "votes": 2, "replies": [ { "id": 2576362, "postDate": "2023-12-27T16:50:12.083Z", "content": "As you are dealing with 3d images, you can view them at different sides, so there are XY, ZX and ZY. Think of that as a cube which you can look at from different angles.
", "rawMarkdown": "As you are dealing with 3d images, you can view them at different sides, so there are XY, ZX and ZY. Think of that as a cube which you can look at from different angles.", "votes": 3 }, { "id": 2576585, "postDate": "2023-12-27T22:33:40.620Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa096d4423afdf74a85957f917ed604f5%2FSelection_999(4450).png?generation=1703716412990283&alt=media\")
more infor
", "rawMarkdown": ".png?generation=1703716412990283&alt=media)\n\nmore infor", "votes": 3 } ] }, { "id": 2582421, "postDate": "2024-01-01T15:19:38.600Z", "content": "did you try using a combination of these strats, or did you get excellent results trying each of them individually?
", "rawMarkdown": "did you try using a combination of these strats, or did you get excellent results trying each of them individually?" } ] }, { "id": 2565193, "postDate": "2023-12-17T18:47:04.473Z", "content": "correct way to do validation:
\nNOTE:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F1c0c51894ac5dfa54adf769e3819685a%2FSelection_999(4439).png?generation=1702839533752331&alt=media\")
another tip:
\nmy previous two winning kaggle solution has huge shake up. I could win because i kept to the follwing principles. instead of the best parameter (e.g. threshold) for the solution:
in summary, in addition to best results, think of ways to
\nhttps://www.kaggle.com/code/junkoda/fast-surface-dice-computation
\nbased on the fast surface dice computation by @junkoda, i did an extensive study on the effect of threshold and the local lb metric. For each model after the training epoch, i made measurements.
\nconclusion:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F16d9246986625481df350c42a67d8e98%2FSelection_999(4341).png?generation=1702270136436198&alt=media\")
I'm using a combination of 0.5Dice loss and 0.5BCE loss. The loss is continuously decreasing, but it seems that the surface dice and LB are also increasing. Perhaps Dice loss is more suitable?
", "rawMarkdown": "I'm using a combination of 0.5*Dice loss and 0.5*BCE loss. The loss is continuously decreasing, but it seems that the surface dice and LB are also increasing. Perhaps Dice loss is more suitable?", "votes": 1 }, { "id": 2556937, "postDate": "2023-12-11T05:02:50.523Z", "content": "Is the model trained on xy xz yz or only inferred on xy xz yz?
", "rawMarkdown": "Is the model trained on xy xz yz or only inferred on xy xz yz?", "votes": 1, "replies": [ { "id": 2556969, "postDate": "2023-12-11T05:41:15.430Z", "content": "model is trained in xy xz yz
", "rawMarkdown": "model is trained in xy xz yz ", "votes": 4 } ] }, { "id": 2562703, "postDate": "2023-12-15T16:01:34.850Z", "content": "Hello, may I ask how you do multi-view training during the training process, is it cycling three axes training in one epoch? I put the data of the three views into a dataset, set the batchsize to 1, and shuffle it, but it doesn't seem to work well. Could you please make a reply? Thank you very much
", "rawMarkdown": "Hello, may I ask how you do multi-view training during the training process, is it cycling three axes training in one epoch? I put the data of the three views into a dataset, set the batchsize to 1, and shuffle it, but it doesn't seem to work well. Could you please make a reply? Thank you very much", "votes": 2, "replies": [ { "id": 2562785, "postDate": "2023-12-15T16:46:53.713Z", "content": " ():\n train_id=[]\n n name:\n d = meta_data[n]\n D,H,W = d.image.shape\n train_id += [ (d.name, i, ) i (H) ]\n train_id += [ (d.name, i, ) i (D) ]\n train_id += [ (d.name, i, ) i (W) ]\n train_id\n\ntrain_id = make_train_id(...)\n\n\n ():\n ():\n self.sample_id = sample_id\n self.augment = augment\n self.length = (self.sample_id)\n\n unique_name=[]\n name,i,axis sample_id:\n name unique_name:\n unique_name.append(name)\n self.unique_name=(unique_name)\n\n ():\n string = \n string += \n string += \n string\n\n ():\n self.length\n\n ():\n name,i,axis = self.sample_id[index]\n\n d = DATA_META[name]\n D,H,W = d.image.shape\n\n axis == :\n image = d.image[i]\n vessel = d.vessel[i]\n axis == :\n image = d.image[:, i]\n vessel = d.vessel[:, i]\n axis == :\n image = d.image[:, :, i]\n vessel = d.vessel[:, :, i]\n\n image = np.ascontiguousarray(image)\n vessel = np.ascontiguousarray(vessel)\n\n self.augment :\n image, vessel = self.augment(image, vessel)\n\n image = np.ascontiguousarray(image)\n vessel = np.ascontiguousarray(vessel)\n \n\n r = {}\n r[] = index\n r[ ] = (name,i,axis)\n r[ ] = torch.from_numpy(image).()\n r[] = torch.from_numpy(vessel).()\n r\nThis seems similar to doing a three-view dataset offline, whether you used a batchsize of 1 or padded the different viewing images to the same size.
", "rawMarkdown": "This seems similar to doing a three-view dataset offline, whether you used a batchsize of 1 or padded the different viewing images to the same size.", "votes": 1 } ] } ] }, { "id": 2555308, "postDate": "2023-12-09T20:32:34.497Z", "content": "example of simple 3d flood fill
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F56ca19ca89d0f127a7e210ec6700bc3e%2FPeek%202023-12-10%2004-15.gif?generation=1702153952547969&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa4fc9bf7de3d501b6aa9b03373731039%2FSelection_999(4312).png?generation=1702153942671435&alt=media\")
iteration in range():\n \n t in range(,,-): \n (t)\n = dense[t+]\n = image[t+]\n = image[t]\n\n = curr_image[:-,:-]\n = prev_mask*prev_image\n\n =\n = \n += np.abs(m-prev[:-,:-])<th #[ , ]\n += np.abs(m-prev[: ,:-])<th #[ , ]\n += np.abs(m-prev[:-,:-])<th #[-, ]\n += np.abs(m-prev[:-,:-])<th #[ ,-]\n += np.abs(m-prev[: ,:-])<th #[ ,-]\n += np.abs(m-prev[:-,: ])<th #[-,-]\n += np.abs(m-prev[:-,: ])<th #[ , ]\n += np.abs(m-prev[: ,: ])<th #[ , ]\n += np.abs(m-prev[:-,: ])<th #[-, ]\n\n = diff>\n \n\n += grow\n \n ('predict',(predict>).astype(np.float32))\n .waitKey()\nIMPORTANT!!!!!
\nThis is the black magic and the trick to winning!!!!
results of 3d flood fill with seed = one pixel
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fdc2eabf5b33f2f89c160fdeb94532138%2FSelection_999(4168).png?generation=1701347716526450&alt=media\")
this reminds me of 3d SAM (segment anything model)
\nif so, we can use self-supervised learning to train the encoder
the seed will be prompt.
\nthe decoder is to perform floodfill segmentation
", "rawMarkdown": "this reminds me of 3d SAM (segment anything model)\nif so, we can use self-supervised learning to train the encoder \n\nthe seed will be prompt. \n \nthe decoder is to perform floodfill segmentation ", "votes": 2, "replies": [ { "id": 2544343, "postDate": "2023-11-30T19:01:08.027Z", "content": "volumetric rendering of image volume
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F4b30a20e48d5434eb5d9e3263c7591fd%2FPeek%202023-12-01%2002-57.gif?generation=1701370799773034&alt=media\")
it shows the uniform intensity of walls of the vessel. that is why flood fill works well
", "rawMarkdown": "volumetric rendering of image volume\n\n\n\n\nit shows the uniform intensity of walls of the vessel. that is why flood fill works well", "votes": 1 }, { "id": 2544387, "postDate": "2023-11-30T19:32:36.910Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F00e9ea02daf6924a01066d9704166d59%2FSelection_999(4171).png?generation=1701372684763901&alt=media\")
this solves the puzzle why there are so much noise in prediction.
\nwe need some median filtering layer?
\nhttps://github.com/llmpass/medianDenoise
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F6f8c6f09bb13beef8496958bca3a2a97%2FSelection_999(4186).png?generation=1701419507768841&alt=media\")
hyptersis thresholding prediction as an alterantive to flood filling on input
", "rawMarkdown": ".png?generation=1701419507768841&alt=media)\n\nhyptersis thresholding prediction as an alterantive to flood filling on input", "votes": 3 }, { "id": 2547016, "postDate": "2023-12-03T05:16:38.150Z", "content": "nice rendering
\nhttps://www.youtube.com/watch?v=wZbfNJxTxqc
wrong annotation !!!!!
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ff8f783cd4318a8853e5d628435131879%2FSelection_999(4156).png?generation=1701314688469717&alt=media\")
i think there is offset bug in annotion for kidney2
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fc736ce3630a4beae36d445ef2e97a962%2FSelection_999(4159).png?generation=1701317221007709&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ff09a93d56aeca55fd5a4e253f3e9f919%2FSelection_999(4158).png?generation=1701317230860236&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fc7cd3b6255a02c133a72e4c8ba7afb84%2FSelection_999(4157).png?generation=1701317239646900&alt=media\")
shift error happens in hidden test datat ground truth before in kaggle competitions…
\nanyone want to probe that?
only top half of kidney 2 annotation are shifted
\nTOP:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fbe2a7252e15844c7b88eb3db63a324b6%2FSelection_999(4437).png?generation=1702748975416857&alt=media\")
BOTTOM:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fe33a6f4a697a929c884f53d4d74e6200%2FSelection_999(4435).png?generation=1702748999529614&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fe3425d8daff24d619bd70dcbd5704ba4%2FSelection_999(4214).png?generation=1701580587953427&alt=media\")
pl = pv.\n\n mhit = pv.).T).glyph(geom=pv.)\n mfp = pv.).T).glyph(geom=pv.)\n mmiss = pv.).T).glyph(geom=pv.)\n pl.add\n pl.add\n pl.add\n pl.show\nmy estimate at one week before the competition:
\nfinal top-1 private leaderboard: private lb 0.890 to 0.880
\ngold cut-off : private lb 0.875 to 0.880
It will be interesting to see how it plays out. I am looking forward to people's write ups so I can understand how they go from 0.8 -> ~0.86 on the leaderboard. I've struggled to get my 2D models (based on SeResNet) to score higher than 0.8. I have learned a lot, but there is plenty more to learn.
", "rawMarkdown": "It will be interesting to see how it plays out. I am looking forward to people's write ups so I can understand how they go from 0.8 -> ~0.86 on the leaderboard. I've struggled to get my 2D models (based on SeResNet) to score higher than 0.8. I have learned a lot, but there is plenty more to learn.", "replies": [ { "id": 2623007, "postDate": "2024-01-27T21:02:10.803Z", "content": "Actually, I have struggled to get train and validation dice scores over 0.84 (kidney 1 train, kidney 3 val) without overfitting
", "rawMarkdown": "Actually, I have struggled to get train and validation dice scores over 0.84 (kidney 1 train, kidney 3 val) without overfitting" }, { "id": 2623113, "postDate": "2024-01-28T01:36:22.463Z", "content": "\"so I can understand how they go from 0.8 -> ~0.86 on the leaderboard. \"
\n\"I have struggled to get train and validation dice … without overfitting\"
answer to both : train augmentation.
\ngenerally(i.e. this and other competitions) all image train augmentation can be classified as
\nsome make the results very worst, some will improve significantly
\nvisualise, visualise, visualise
\nThank you @hengck23 for always being so helpful on these forums! I noticed that the train augmentation helps improve the consistency between the train, test, and even leaderboard. Perhaps I need to try more! I appreciate the advice to inspect early epochs. I hadn't done that, I typically just save the best weights and inspect the result. I'll give it a try :).
", "rawMarkdown": "Thank you @hengck23 for always being so helpful on these forums! I noticed that the train augmentation helps improve the consistency between the train, test, and even leaderboard. Perhaps I need to try more! I appreciate the advice to inspect early epochs. I hadn't done that, I typically just save the best weights and inspect the result. I'll give it a try :). " }, { "id": 2626979, "postDate": "2024-01-30T12:08:10.753Z", "content": "visualisation for the first 0.25 epoch
\n![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F425a8cc975f235331a07db0a06f88ad1%2FSelection_999(4713).png?generation=1706616477661913&alt=media\")
\n![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fc5236e5a84bfd183891f8a229094914f%2FSelection_999(4714).png?generation=1706616489116957&alt=media\")
Very interesting! I've been working on doing similar after your comment. Your advice has been very helpful, I've been implementing it and hoping it can help me move up on the leaderboard before the final :)
", "rawMarkdown": "Very interesting! I've been working on doing similar after your comment. Your advice has been very helpful, I've been implementing it and hoping it can help me move up on the leaderboard before the final :)" } ] }, { "id": 2623085, "postDate": "2024-01-28T00:27:55.863Z", "content": "This one has been a heart breaker, I have been stuck with my current score for about a month now and it seems unless I come up with something I will be just shy of gold. I am interested to see what others did because I feel like I have tried so many things that just don’t seem to help.
", "rawMarkdown": "This one has been a heart breaker, I have been stuck with my current score for about a month now and it seems unless I come up with something I will be just shy of gold. I am interested to see what others did because I feel like I have tried so many things that just don’t seem to help.", "votes": 2, "replies": [ { "id": 2623137, "postDate": "2024-01-28T02:15:48.037Z", "content": "I understand. I got really excited about a custom loss function I found in the literature and for my models below 0.8 on the public leaderboard it always seemed to improve the score, but when I tried it with the model that scored well it didn't improve. I keep hitting a wall. Still, it's been a fun competition! I've never worked on a proper segmentation problem before and I've learned a lot! Good luck getting into gold as the leaderboard shifts with the private data.
", "rawMarkdown": "I understand. I got really excited about a custom loss function I found in the literature and for my models below 0.8 on the public leaderboard it always seemed to improve the score, but when I tried it with the model that scored well it didn't improve. I keep hitting a wall. Still, it's been a fun competition! I've never worked on a proper segmentation problem before and I've learned a lot! Good luck getting into gold as the leaderboard shifts with the private data.", "votes": 1 }, { "id": 2623175, "postDate": "2024-01-28T03:23:09.900Z", "content": "Thank you! Good luck to you too! Yes the learning is what is important and I have definitely done a lot of that! Good to get reminders every once in awhile!
", "rawMarkdown": "Thank you! Good luck to you too! Yes the learning is what is important and I have definitely done a lot of that! Good to get reminders every once in awhile!", "votes": 1 } ] }, { "id": 2625364, "postDate": "2024-01-29T11:09:57.290Z", "content": "one more information. from the paper: if train = kidney1 upper, valid = kidney1 lower, surface dice is 0.95.
\nBut i think that is not using TTA.
\n(i haven't done experiment on upper/lower split).
I estimate with TTA and better processing, surface dice is 0.96+.
\nThis means that under perfect data (i.e. no domain shift), expect a 0.03+
", "rawMarkdown": "one more information. from the paper: if train = kidney1 upper, valid = kidney1 lower, surface dice is 0.95.\nBut i think that is not using TTA.\n(i haven't done experiment on upper/lower split).\n\nI estimate with TTA and better processing, surface dice is 0.96+.\n\nThis means that under perfect data (i.e. no domain shift), expect a 0.03+", "replies": [ { "id": 2625675, "postDate": "2024-01-29T14:38:28.967Z", "content": "I noticed something today which you might find interesting.
\nFor my validation set, which is kidney 3 dense, I generated a submission.csv using all the images.
\nThen I broke it into 10 samples to calculate the surface dice score, each sample with a step size of 10, such that my first sample would consider img 0, 10, 20 etc., the second would consider img 1, 11, 21 and so on.
\nEach individual sample scored above 0.94, and the avg was over .94 as well. However, when I calculated dice score over the full CV set without breaking it into sub samples, the dice score became 0.86. This has been confusing me. Im using @Jun Koda's method. https://www.kaggle.com/datasets/junkoda/sennet-score.
\nObviously my actual score on the public test set is around .835, which is much closer to .86 than .94.
\nI thought that the surface dice scores are being averaged over the sample, but that doesn't seem to be the case.
\nDo you have any thoughts?
What you are doing does not make sense -> the competition metric is computed over consecutive frames: the dice score is computed on the surface area defined by two consecutive frames. So if you compute the score with images separated by 10 images this will create a very poor ground truth for the surface area.
\nIf you wish to see you scores on subsets, you should take adjacent frames like 0 to 100, then 100 to 200 etc… The local scores should then average to the global score.
", "rawMarkdown": "What you are doing does not make sense -> the competition metric is computed over consecutive frames: the dice score is computed on the surface area defined by two consecutive frames. So if you compute the score with images separated by 10 images this will create a very poor ground truth for the surface area.\n\nIf you wish to see you scores on subsets, you should take adjacent frames like 0 to 100, then 100 to 200 etc... The local scores should then average to the global score.", "votes": 2 }, { "id": 2625774, "postDate": "2024-01-29T15:26:32.107Z", "content": "Wow, that clarifies so much for me. Thank you very much. I suppose dice loss is very different.
\nI have learned a valuable lesson.
i wonder did anyone try this:
\nThat's a really good idea, will try and let everyone know the results here
", "rawMarkdown": "That's a really good idea, will try and let everyone know the results here", "votes": 3 }, { "id": 2626640, "postDate": "2024-01-30T06:33:05.297Z", "content": "from the paper: if train = kidney1 upper, valid = kidney1 lower, surface dice is 0.95.
\nCan you tell me about this paper?
\nRelabeling \"kidney3 sparse\" will help, but I don't think it's easy to get a perfect model by using k1 and k2 for validation in \"kidney3 dense\" training. I'm very curious about how much validation data was used for training in the above paper, including the specific training model.
", "rawMarkdown": "from the paper: if train = kidney1 upper, valid = kidney1 lower, surface dice is 0.95.\n\nCan you tell me about this paper?\n\nRelabeling \"kidney3 sparse\" will help, but I don't think it's easy to get a perfect model by using k1 and k2 for validation in \"kidney3 dense\" training. I'm very curious about how much validation data was used for training in the above paper, including the specific training model." } ] } ] }, { "id": 2589018, "postDate": "2024-01-06T01:16:46.003Z", "content": "wrong strategy?
\ni have been trying detect small vessels improve LB score becuase stated paper this one problem.\nBut i realize this may be a good strategy?\n\ni use train= kidney1 dense + kidney3 dense\nvalidation= kidney2 sparse ( slice >= avoid annotation shift )\n\ni note recent popular public kernel based seresnext50_32x4d dunet which can easily object lb . seresnext50_32x4d has imagenet top1 range ~. My experiments shows seresnext50_32x4d dunet has low hit rate (>), has very low fp rate (~) well. i am surprise can lb single model.\n\ni have been using nextVIT base transformer, imagenet top1 ~, lb . It has very high hitrate (>) low fp rate (~).\n\n turns out lb based surface pixels, small vessel has less surface ( contribute little lb score) !!!\ne.g. kidney3 sparse (% annotation) has surface dice compared kidney3 dense\n\n other , wrong choice metric host interest detect small vessel\nupdate on 06-JAN
\ntrick to get lb 0.87+
make a base model in the range of 0.864 (e.g. either seresnext50_32x4d or nextVIT base 2d unet)
ensemble of the two will give 0.87+
ensemble improve lb score:
Thank you for sharing. Different models might have different thresholds. How do you perform model fusion? Do you average the probabilities and then use a compromise threshold?
", "rawMarkdown": "Thank you for sharing. Different models might have different thresholds. How do you perform model fusion? Do you average the probabilities and then use a compromise threshold?", "votes": 2 }, { "id": 2589337, "postDate": "2024-01-06T09:44:17.340Z", "content": "Thank you for sharing. I would like to know approximately how long it takes for a single model of amp inference, and would it be faster to use TensorRT instead
", "rawMarkdown": "Thank you for sharing. I would like to know approximately how long it takes for a single model of amp inference, and would it be faster to use TensorRT instead", "votes": 1 }, { "id": 2589420, "postDate": "2024-01-06T11:20:49.200Z", "content": "seresnext50_32x4d : 2hr +
\nnextVIT base 2d : 3hr +
i haven't use tensorRT yet
", "rawMarkdown": "seresnext50_32x4d : 2hr +\nnextVIT base 2d : 3hr +\n\ni haven't use tensorRT yet", "votes": 1 }, { "id": 2589498, "postDate": "2024-01-06T12:40:51.307Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fb2411a886d57c613e767a0aaff915faf%2FSelection_999(4532).png?generation=1704544815068725&alt=media\")
i have not decided. but the threshold is smiliar for same validation set even for different model
", "rawMarkdown": ".png?generation=1704544815068725&alt=media)\n\n@lihaoweicvch \n\ni have not decided. but the threshold is smiliar for same validation set even for different model", "votes": 4 }, { "id": 2589696, "postDate": "2024-01-06T15:17:07.227Z", "content": "Thank you for answer.Is this the duration of reasoning for only one perspective without TTA? When I take TTA from three perspectives, seresnext50_ 32x4d took approximately 6 hours to reason
", "rawMarkdown": "Thank you for answer.Is this the duration of reasoning for only one perspective without TTA? When I take TTA from three perspectives, seresnext50_ 32x4d took approximately 6 hours to reason", "votes": 1 }, { "id": 2589706, "postDate": "2024-01-06T15:25:31.097Z", "content": "time is for infer with the usual xy,zx,zy + 5 TTA
", "rawMarkdown": "time is for infer with the usual xy,zx,zy + 5 TTA", "votes": 1 } ] }, { "id": 2590673, "postDate": "2024-01-07T11:19:37.070Z", "content": "@hengck23 How do you compute your FPR ? Your definition probably differs from the one I'm used to regarding normalization.
\nDo you use (pred * (1 - truth)).sum() / truth.sum() ?
\nSame for TPR, you are using (pred * truth).sum() / truth.sum(), right ?
Thanks!
", "rawMarkdown": "@hengck23 How do you compute your FPR ? Your definition probably differs from the one I'm used to regarding normalization.\n\nDo you use `(pred * (1 - truth)).sum() / truth.sum()` ?\nSame for TPR, you are using `(pred * truth).sum() / truth.sum()`, right ?\n\nThanks!\n", "votes": 1, "replies": [ { "id": 2590803, "postDate": "2024-01-07T12:53:54.400Z", "content": "def np_metric(, truth):\n p = (>0.5)\n t = (truth>0.5)\n hit = (p*t).()\n fp = (p*(1-t)).()\n t_sum = t.()\n p_sum = p.()\n hit, fp, t_sum, p_sum\n\n#-------------------------------------\n = >0.5\nhit, fp, t_sum, p_sum = np_metric(, truth)\nhit = hit/t_sum\nfp = fp/p_sum\nprecision and recall are better words
", "rawMarkdown": "```\ndef np_metric(predict, truth):\n\tp = (predict>0.5)\n\tt = (truth>0.5)\n\thit = (p*t).sum()\n\tfp = (p*(1-t)).sum()\n\tt_sum = t.sum()\n\tp_sum = p.sum()\n\treturn hit, fp, t_sum, p_sum\n\n#-------------------------------------\npredict = prob>0.5\nhit, fp, t_sum, p_sum = np_metric(predict, truth)\nhit = hit/t_sum\nfp = fp/p_sum\n\n```\n\nprecision and recall are better words", "votes": 4 }, { "id": 2590930, "postDate": "2024-01-07T14:46:00.647Z", "content": "Gotcha, thanks.
\nFYI, those are my scores on kidney_2:
\n\nhit: 0.9549\nfp: 0.0507\nCV: 0.84\nLB ~0.80\n
\nMy LB scores are lagging behind compared to yours but my CV scores are close. Not sure if it is because I have too many FPs or not.
submit with different thresholds.
\nCV and LB correlations are weak. but i do see a general linear relationship: lb = cv + bias
\n(cv and lb don't share the best optimal threshold, which is a** problem for us to avoid shakeup**)
\"FYI, those are my scores on kidney_2:\"
\nThere are annotation errors in kidney 2. so your lb score should be better if your threshold is correct
", "rawMarkdown": "submit with different thresholds.\n\nCV and LB correlations are weak. but i do see a general linear relationship: lb = cv + bias\n(cv and lb don't share the best optimal threshold, which is a** problem for us to avoid shakeup**)\n\n\"FYI, those are my scores on kidney_2:\"\n\nThere are annotation errors in kidney 2. so your lb score should be better if your threshold is correct", "votes": 2 }, { "id": 2591184, "postDate": "2024-01-07T17:50:56.710Z", "content": "@theoviel
\ni think your metric are for model without TTA.
you need to do TTA (it is common that 2d Unet can completely miss vessel in one view and detect in another). TTA may not improve CV, but it should improve LB.
", "rawMarkdown": "@theoviel \ni think your metric are for model without TTA.\n\nyou need to do TTA (it is common that 2d Unet can completely miss vessel in one view and detect in another). TTA may not improve CV, but it should improve LB.", "votes": 3 } ] } ] }, { "id": 2554978, "postDate": "2023-12-09T15:30:08.363Z", "content": "trick to get lb 0.835
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F7f8c7dca190065f9d3b98f880ef7cd21%2FSelection_999(4310).png?generation=1702135804600547&alt=media\")
def forward(self, batch):\n x = batch['image']\n x = x.(-1,3,-1,-1) \n B, C, , W = x.shape\n\n = []\n xx = self.stem0(x); .(xx)\n xx = F.avg_pool2d(xx,kernel_size=2,stride=2)\n xx = self.stem1(xx); .(xx)\n\n = self.encoder #convnext\n x = .(x);\n\n x = .stages[0](x); .(x)\n x = .stages[1](x); .(x)\n x = .stages[2](x); .(x)\n x = .stages[3](x); .(x)\n ##[(f'encode_{i}', .shape) i, enumerate()]\n\n last, = self.decoder(\n feature=[-1], skip=[:-1][::-1]\n )\n ##[(f'decode_{i}', .shape) i, enumerate()]\n ##('last', last.shape)\n\n vessel = self.vessel(last)\nyou can go to super-resolution
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F86402d134c124fd64c7695762e052762%2FSelection_999(4311).png?generation=1702136734642968&alt=media\")
HINT:
\nalternatively, you can train a real super-resolution net (you have 20um imags from the HIP-CT website)
local CV/LB
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa034d0e1646416700a5b6fadd8e26bbb%2FSelection_999(4315).png?generation=1702169707335185&alt=media\")
lb0.829 …
\nnormalisation is the key
\n norm_by_percentile(volume, low=, high=., alpha=.):\n = np.percentile(volume,low)\n = np.percentile(volume,high)\n = (volume-xmin)/(xmax-xmin)\n :\n [x>]=(x[x>]-)*alpha +\n [x<]=(x[x<])*alpha\n \n x\nnormalised by volume/subvolume, not image
", "rawMarkdown": "lb0.829 ...\n\nnormalisation is the key\n\n```\ndef norm_by_percentile(volume, low=10, high=99.8, alpha=0.01):\n\txmin = np.percentile(volume,low)\n\txmax = np.percentile(volume,high)\n\tx = (volume-xmin)/(xmax-xmin)\n\tif 1:\n\t\tx[x>1]=(x[x>1]-1)*alpha +1\n\t\tx[x<0]=(x[x<0])*alpha\n\t#x = np.clip(x,0,1)\n\treturn x\n\n```\n\nnormalised by volume/subvolume, not image\n", "votes": 6, "replies": [ { "id": 2554394, "postDate": "2023-12-09T05:50:07.933Z", "content": "@hengck23 - is this the right way to read 3D volume because i get negative stride error
\n (torch.utils.data.Dataset):\n ():\n self.img_voxel = create_3d_voxel(img_voxel_paths) \n self.msk_voxel = create_3d_voxel_msk(msk_voxel_paths) msk_voxel_paths \n self.transforms = transforms\n\n ():\n self.img_voxel.shape[] \n\n ():\n img = self.img_voxel[index, :, :].copy() \n img = np.expand_dims(img, axis=) \n\n self.msk_voxel :\n msk = self.msk_voxel[index, :, :].copy() \n msk = np.expand_dims(msk, axis=) \n\n self.transforms:\n data = self.transforms(image=img, mask=msk)\n img = data[]\n msk = data[]\n\n torch.tensor(img, dtype=torch.float32), torch.tensor(msk, dtype=torch.float32)\n :\n self.transforms:\n data = self.transforms(image=img)\n img = data[]\n\n torch.tensor(img, dtype=torch.float32)\ndef do_random_flip_rotate(, vessel):\n ..rand()<:\n = .flip(, axis=) #horizontal\n vessel = .flip(vessel,axis=)\n ..rand()<:\n = .flip(, axis=)\n vessel = .flip(vessel,axis=)\n ..rand()<:\n k = ..choice([,,])\n = .rot90(, k, =[,])\n vessel = .rot90(vessel,k, =[,])\n\n = .ascontiguousarray()\n vessel = .ascontiguousarray(vessel)\n , vessel\nstill i face the same error
", "rawMarkdown": "still i face the same error " } ] } ] }, { "id": 2541221, "postDate": "2023-11-28T09:58:56.980Z", "content": "Yellow: MISS
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F88f268f9b529a2f68e9e111d53662121%2FPeek%202023-11-28%2017-51.gif?generation=1701165525275937&alt=media\")
\nGreen: FP
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F5cc5e3d3dcdca56155b2f4fa5b1ed96e%2FPeek%202023-11-28%2017-53.gif?generation=1701165506289118&alt=media\")
\nRed: HIT
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F57447a54514accc6fd1d3b1cfdebcef6%2FPeek%202023-11-28%2017-55.gif?generation=1701165488221412&alt=media\")
\nALL
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F65417fbd421b7741d417ae9ac74ee5ab%2FPeek%202023-11-28%2017-56.gif?generation=1701165472038050&alt=media\")
beware !!!
\nit is not any vessel, but those from \" arterial vascular tree \"
data page: \"kidney_1_dense - The whole of a right kidney at 50um resolution. The entire 3D arterial vascular tree … \"
\nhere you can see that not all vessel are annotated (kidney dense 1)
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F36a823d5c809fe74c99f774a7cfbd507%2FSelection_999(4051).png?generation=1700964826477601&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F0a6889dcdc38de11f00ae61ae74b68f1%2FSelection_999(4050).png?generation=1700964841203613&alt=media\")
i was wondering do we need to separate arterial and venous?
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F622be63c31bba65e4de644cad69ae98f%2FSelection_999(4052).png?generation=1700965019077924&alt=media\")
i am surprised that the deep CNN nework can somhow differentate between venous and arterial (target) vessel from one imgae.
\ncan anyone explain why?
\n(e.g. they have different apperance due to dye/contrast when the Hip CT scan is taken ??? or arterial vessel have thicker wall ???)
The structure of the membrane differs significantly between arteries and veins. Arteries are richer in elastic fibres and thicker than veins. The walls of arteries and veins are very different, as can be seen in the linked histological image (HE-stain).
\nhttps://www.kidneypathology.com/English_version/Vessels_histology.htm
@yosukeyama
\n\"The walls of arteries and veins are very different, \"
Thanks.
\nThat is good news
\nso it seen that we need not grow the tree from scratch. just pure detection (+ post processing) is probably enough.
very bad news !!!
\ni try to do some probing.
public lb set: kidney 5
\nprivate lb set: kidney 6
you may want to verify yourself ( just in case my code is buggy)
", "rawMarkdown": "**very bad news !!!**\ni try to do some probing.\n\npublic lb set: kidney 5\nprivate lb set: kidney 6\n\ndon't overfit ... there could be shakeup. worst still, what if kidney 5 and 6 are of different um resolution???\n---\n\nyou may want to verify yourself ( just in case my code is buggy)", "votes": 6, "replies": [ { "id": 2537624, "postDate": "2023-11-25T11:05:56.733Z", "content": "I think your guess is correct because I had the same result when I did LB Probing before. (Public is calculated on kidney_5 only, Private is calculated on kidney_6 only.)
\nI think it was reasonable to separate Public and Private by kidney_5 and kidney_6 because it is difficult to calculate metrics with such a high computational load more times than necessary.
I agree with you that the information on the test dataset is rarely publicly available, which causes Big Shake in the worst case scenario.
", "rawMarkdown": "I think your guess is correct because I had the same result when I did LB Probing before. (Public is calculated on kidney_5 only, Private is calculated on kidney_6 only.)\nI think it was reasonable to separate Public and Private by kidney_5 and kidney_6 because it is difficult to calculate metrics with such a high computational load more times than necessary.\n\nI agree with you that the information on the test dataset is rarely publicly available, which causes Big Shake in the worst case scenario.", "votes": 3, "replies": [ { "id": 2537870, "postDate": "2023-11-25T15:06:39.347Z", "content": "It was not hard to imagine that this kind of data split is being used.
\nIf the resolution is different, I believe it essentially becomes a completely different image. Since we only have two types of resolution available, appropriate validation is very difficult. While it's not entirely unfeasible, there is a concern that it will turn into a competition heavily influenced by luck. In this case, the top solutions may not necessarily be the superior solutions…
", "rawMarkdown": "It was not hard to imagine that this kind of data split is being used.\n\nIf the resolution is different, I believe it essentially becomes a completely different image. Since we only have two types of resolution available, appropriate validation is very difficult. While it's not entirely unfeasible, there is a concern that it will turn into a competition heavily influenced by luck. In this case, the top solutions may not necessarily be the superior solutions...", "votes": 1 }, { "id": 2547015, "postDate": "2023-12-03T05:10:06.350Z", "content": "what if private kidney6 (500 images) is actually like kidney1 voi set at about 5.2um resolution.
\ni cannot think of a reason why kidney1 voi set is included in the train set ….
Something funny is my best score 0.62 is from mixing everything… I will probably remove the sparse and see if that makes a difference
", "rawMarkdown": "Something funny is my best score 0.62 is from mixing everything... I will probably remove the sparse and see if that makes a difference" } ] } ] }, { "id": 2612593, "postDate": "2024-01-21T13:42:56.723Z", "content": "experiments on scale … yet another proof that it is FP that matters …
\n![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F689d2443d12433d50098972a223f2516%2FSelection_999(4657).png?generation=1705844574982811&alt=media\")
![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F97a9b38d338e4373bd8c7463fcaca5bc%2FSelection_999(4658).png?generation=1705844565213916&alt=media\")
OOPS!!! there is a metric bug
\n(this is behaviour for kidney2. maybe different behaviour for different kidney)
![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F612b59de52eaa1179ccb4c3b16394857%2FSelection_999(4662).png?generation=1705846447352595&alt=media\")
Hello @hengck23,
\nSince your conclusion is that fp matters, have you considered using the TverskyLoss ?
With it you can weight the impact of FP and FN, making the loss closer to the metric. What do you think ?
", "rawMarkdown": "Hello @hengck23,\nSince your conclusion is that fp matters, have you considered using the [TverskyLoss](https://smp.readthedocs.io/en/latest/losses.html#segmentation_models_pytorch.losses.TverskyLoss) ? \n\nWith it you can weight the impact of FP and FN, making the loss closer to the metric. What do you think ?", "votes": 1 } ] } ] }, { "id": 2589059, "postDate": "2024-01-06T03:02:02.963Z", "content": "an interesting idea
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fc777e469f711c48ae77c01ff5632bc8b%2FSelection_999(4527).png?generation=1704510120822206&alt=media\")
How long does it take for you to inference with Transformer, it seems that P100 without fp16 inference is very slow for Transformer, do you have any suggestions for this?
", "rawMarkdown": "How long does it take for you to inference with Transformer, it seems that P100 without fp16 inference is very slow for Transformer, do you have any suggestions for this?", "votes": 3, "replies": [ { "id": 2568996, "postDate": "2023-12-21T02:45:12.713Z", "content": "you can google for fast transformer. i use nextvit, which was used in kaggle rnsa breast mamnographs before. it takes 3to4 hr and can be speedup further by tensort rt to 3 hr
", "rawMarkdown": "you can google for fast transformer. i use nextvit, which was used in kaggle rnsa breast mamnographs before. it takes 3to4 hr and can be speedup further by tensort rt to 3 hr", "votes": 2, "replies": [ { "id": 2569002, "postDate": "2023-12-21T02:53:01.587Z", "content": "Thank you so much!
", "rawMarkdown": "Thank you so much! " }, { "id": 2569013, "postDate": "2023-12-21T03:04:22.723Z", "content": "update: nextVIT-base can get LB0.682 3hr 20 min (just complete in few minutes along so i can get the timing in minutes).
\nit uses 5xTTA + 3 axis (xy,yz,xz) in P100
", "rawMarkdown": "update: nextVIT-base can get LB0.682 3hr 20 min (just complete in few minutes along so i can get the timing in minutes).\n\nit uses 5xTTA + 3 axis (xy,yz,xz) in P100", "votes": 2 }, { "id": 2593750, "postDate": "2024-01-09T12:30:04.533Z", "content": "Seems that nextVIT is a kind of detection model. Should I modify the configuration of Q1&Q2 Net for segmentation task when using? Thanks!
", "rawMarkdown": "Seems that nextVIT is a kind of detection model. Should I modify the configuration of Q1&Q2 Net for segmentation task when using? Thanks!" } ] }, { "id": 2569004, "postDate": "2023-12-21T02:55:23.350Z", "content": "Try fp16 + 2xT4 + nn.DataParallel, for me it's faster than using single P100
\nThe model I choose is very slow [maxvit-small],
\nand scoring single checkpoint with xy/yz/xz + 5tta takes 8 hours…
should try Next-ViT I think :D
", "rawMarkdown": "Try fp16 + 2xT4 + nn.DataParallel, for me it's faster than using single P100\n\nThe model I choose is very slow [maxvit-small], \nand scoring single checkpoint with xy/yz/xz + 5tta takes 8 hours... \n\nshould try Next-ViT I think :D ", "replies": [ { "id": 2569034, "postDate": "2023-12-21T03:23:30.103Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F2d5e11a186a047d2f13d10b613c413a7%2FSelection_999(4447).png?generation=1703128978938372&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F859e0cdb0b7ab614c159c4721cbc7623%2FSelection_999(4446).png?generation=1703128988887048&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F22550ca323f59dec3a1069decd0042db%2FSelection_999(4445).png?generation=1703129001328081&alt=media\")
the fastest i can find
", "rawMarkdown": ".png?generation=1703128978938372&alt=media)\n\n.png?generation=1703128988887048&alt=media)\n\n.png?generation=1703129001328081&alt=media)\n\nthe fastest i can find", "votes": 2 } ] } ] }, { "id": 2583079, "postDate": "2024-01-02T03:59:54.163Z", "content": "did we just find the missing labels (for sparsely annotated ground truth labels)?
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F7d59d1157c40bcddd6013b488ad8adbc%2FSelection_999(4513).png?generation=1704167987606057&alt=media\")
GJ. I believe both sparse and dense have labeling issues (imho, cuz I`m not a doctor). Looked through and found even pretty big vessels missed. Thx for sharing.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F621304%2F6ffbfd469eb507bb65297d3740dc3cba%2Fezgif-1-83993798fe_3.gif?generation=1704212371219171&alt=media\")
yes you are correct.
\nan easy way to prove label error is to check train error after overfitting.
\nnow if the train error cannot get to zero, waht can be the reason?
\nExactly. Heavy backbones perform nearly identically as small ones
", "rawMarkdown": "Exactly. Heavy backbones perform nearly identically as small ones" } ] } ] }, { "id": 2560936, "postDate": "2023-12-14T06:01:20.913Z", "content": "just an idea:
\ntakes 2 nearby slices and generate vessels in between, both image and mask
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F4ee1367bb6b99b88ba2d31c938b15191%2FSelection_999(4402).png?generation=1702533674316871&alt=media\")
i hope there is no bug, but mixing kidney 1 and 3 in train make LB (much worse)
\nreference: train with kidney 1 only: LB 0.757
submit inference mode: add TTA 2xflip,3xrot90 only
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F71837a4178ef96b35c64d94684f90a50%2FSelection_999(4107).png?generation=1701223141188369&alt=media\")
since there is no validation set, i just train to the same number of iterations from previous experiments.
\nthis may not be optimal (and hence not optimal lb)???
I trained initially on kidney 1, by adding also kidney 3 improved my LB from 0.65 -> 0.7. I do not know how to compare their resolutions but looks like they have pretty similar ones with 50um and 50.16um. The question would be if LB is also 50um and private something completly different
", "rawMarkdown": "I trained initially on kidney 1, by adding also kidney 3 improved my LB from 0.65 -> 0.7. I do not know how to compare their resolutions but looks like they have pretty similar ones with 50um and 50.16um. The question would be if LB is also 50um and private something completly different ", "votes": 2, "replies": [ { "id": 2542830, "postDate": "2023-11-29T14:48:01.933Z", "content": "Thanks, i will check again
", "rawMarkdown": "Thanks, i will check again" } ] } ] }, { "id": 2539766, "postDate": "2023-11-27T09:53:46.590Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F9815c2618468fcef78460941c6ec295b%2FSelection_999(4078).png?generation=1701078765760475&alt=media\")
use 3d connected components to filter false positive errors
\nhttps://pypi.org/project/connected-components-3d/
In your notebook https://www.kaggle.com/code/hengck23/lb0-808-resnet50-2d-unet-xy-zy-zx-cc3d you didn't use connected-components-3d:
\ncc_threshold = -1
\nWhat were your results using cc3d?
useful:
\nhttps://github.com/JacobBumgarner/VesselVio/tree/main
\nVesselVio is an open-source application designed for the analysis and visualization of segmented vasculature datasets.
https://blog.research.google/2021/06/a-browsable-petascale-reconstruction-of.html
\nthis competition reminds me of the ancient google paper on flood filling network (FFN)
grow_mask = \n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fac63fb84563f14ea23aa7d6aa4f5f80a%2Fartificialne.jpg?generation=1701075207180931&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F8c6ac1934b1a49c53455c5d6e9059b22%2Fimage4.gif?generation=1701075219778016&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa301038a9c7d5635edc5ded3b75d674d%2FSelection_999(4077).png?generation=1701075493076458&alt=media\")
\nHigh-precision automated reconstruction of neurons with flood-filling networks
\nhttps://www.nature.com/articles/s41592-018-0049-4
\nhttps://www.biorxiv.org/content/10.1101/200675v1.full
\nhttps://www.mpg.de/12130750/neural-networks-connectome
\nhttps://www.youtube.com/watch?v=46SksPonI8I
GCO post processing:
\n(Global Constructive Optimization algorithm for generating smaller vessels)
\n[1] A Hybrid Approach to Full-Scale Reconstruction of Renal Arterial Network
\nhttps://arxiv.org/pdf/2303.01837.pdf
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fd5798930fc9496d763a11c85748f2ed1%2FSelection_999(4060).png?generation=1701038372973913&alt=media\")
(e) is results from your deep network.
\n(j) is the results of tree growing with GCO
proposed solution:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F23665654c97f956393f163ab991e4bb6%2FSelection_999(4059).png?generation=1701037259106090&alt=media\")
are you planning to apply a patch based approach? or taking the whole image into account? will we be able to fit the whole image into memory with the kaggle infrastructure?
", "rawMarkdown": "are you planning to apply a patch based approach? or taking the whole image into account? will we be able to fit the whole image into memory with the kaggle infrastructure? ", "votes": 1 } ] }, { "id": 2560267, "postDate": "2023-12-13T12:29:28.990Z", "content": "useful
\nhttps://github.com/PengyiZhang/RegionGrowth
\nhttps://github.com/ylmzkaan/3DRegionGrowing
\nhttp://notmatthancock.github.io/2017/10/09/region-growing-wrapping-c.html
high resolution 2d/3d unet solution !!!
\ndemo notebook :https://www.kaggle.com/code/hengck23/2d-to-3d-unet-demo
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F1616c648d75df30b4c8ccceedb5eb8f6%2FSelection_999(4362).png?generation=1702314125458912&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F74e23e8d6c4b12748102003e0981e98a%2FSelection_999(4363).png?generation=1702314137160715&alt=media\")
how to properly design 3d solution.
\nif you cannot fit a single 3d net work (because of memory GPU constraint),
\nyou properly need at least 2 net (of different hierarchy scale).
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Faba1ca9b68fe072afcf39f2eb9398793%2FSelection_999(4365).png?generation=1702316114508754&alt=media\")
posting it here for visibility Patch based 3D UNET + TF + TPU: https://www.kaggle.com/code/dhinkris/training-patch-based-3d-unet-tf-tpu
", "rawMarkdown": "posting it here for visibility Patch based 3D UNET + TF + TPU: https://www.kaggle.com/code/dhinkris/training-patch-based-3d-unet-tf-tpu", "votes": 1, "replies": [ { "id": 2557807, "postDate": "2023-12-11T17:15:07.697Z", "content": "you should try these pretrain 3d unet
\nhttps://github.com/Tencent/MedicalNet
@hengck23 how do you stack the 2Dslice to 3D in kaggle inference , i get memory overflow error
", "rawMarkdown": "@hengck23 how do you stack the 2Dslice to 3D in kaggle inference , i get memory overflow error ", "votes": 1, "replies": [ { "id": 2556823, "postDate": "2023-12-11T02:48:57.367Z", "content": "i don't train in kaggle. i use my local machine
", "rawMarkdown": "i don't train in kaggle. i use my local machine", "replies": [ { "id": 2557162, "postDate": "2023-12-11T09:00:58.290Z", "content": "it is during inference, i stack the images as a 3D voxel but it lands in Your notebook tried to allocate more memory than is available
", "rawMarkdown": "it is during inference, i stack the images as a 3D voxel but it lands in Your notebook tried to allocate more memory than is available " }, { "id": 2557166, "postDate": "2023-12-11T09:03:24.763Z", "content": "stack as uint16 or uint8.
\nstack only one at a time (not both kidney 6,5)
ok, 1st stack kidney 5 generate df and stack kidney 6 generate df and then concat them
", "rawMarkdown": "ok, 1st stack kidney 5 generate df and stack kidney 6 generate df and then concat them ", "votes": 1 } ] }, { "id": 2558273, "postDate": "2023-12-12T03:28:12.807Z", "content": "take a look at this thread: https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/453986
", "rawMarkdown": "take a look at this thread: https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/453986", "votes": 1 } ] }, { "id": 2555747, "postDate": "2023-12-10T06:38:29.373Z", "content": "discrepancies happens and takes me several \"wasting\" hours to debug
\nthis affects how we set the threshold …
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ff40d26a2fd0ef81938725dd053bf9dfe%2FSelection_999(4317).png?generation=1702190263074677&alt=media\")
trying augmentation
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F711961849f08998cf722c6a864bfc5b7%2FPeek%202023-12-07%2022-00.gif?generation=1701957714420766&alt=media\")
i think this is a good pretrain/self-supervised or aux loss:
\ninput \npredict \naux target \n\npredict how the volume will \n\ngood idea👍
", "rawMarkdown": "good idea👍", "replies": [ { "id": 2552505, "postDate": "2023-12-07T14:12:46.793Z", "content": "by measuring the aux error of hidden kidney5 and kidney6, we can also know if these can similar to train data or not :)
\nif we have enough time, online fine tuning is possible
treat extrapolate as image prompting problem
\nSequential Modeling Enables Scalable Learning forLarge Vision Models
\nhttps://yutongbai.com/lvm.html
quick experiments for kidney3 sparse as train:
\nresnext26d- single class
\nthreshold=0.25
\nTTA=flip+rot90
surface dice for kidney3 sparse/dense = 0.98
\ntrain = kidney1 dense
\nvalid = kidney3 dense
\nLB = 0.747
train = kidney1 dense + kidney3 sparse
\nvalid = kidney3 dense
\nLB = 0.764
no parameter tuning, etc
", "rawMarkdown": "quick experiments for kidney3 sparse as train:\n\nresnext26d- single class\nthreshold=0.25\nTTA=flip+rot90\n\nsurface dice for kidney3 sparse/dense = 0.98 \n\n---\n\ntrain = kidney1 dense \nvalid = kidney3 dense\nLB = 0.747\n\n---\n\ntrain = kidney1 dense + kidney3 sparse \nvalid = kidney3 dense\nLB = 0.764\n\nno parameter tuning, etc\n", "votes": 1 }, { "id": 2541738, "postDate": "2023-11-28T18:29:58.130Z", "content": "What does \"model trained with multiview\" mean? Does it mean that supppose you have a simple 2d unet model and 3d voxel input, and during training you randomly choose which plane (XY, YZ, XZ) to take from that voxel and take the according mask?
\nDuring inference we can use 3 tta (take 3 different views) for the same pixel, am I correct?
", "rawMarkdown": "What does \"model trained with multiview\" mean? Does it mean that supppose you have a simple 2d unet model and 3d voxel input, and during training you randomly choose which plane (XY, YZ, XZ) to take from that voxel and take the according mask?\n\nDuring inference we can use 3 tta (take 3 different views) for the same pixel, am I correct?", "votes": 1, "replies": [ { "id": 2541739, "postDate": "2023-11-28T18:32:52.837Z", "content": "you can correct
", "rawMarkdown": "you can correct", "votes": 1, "replies": [ { "id": 2541787, "postDate": "2023-11-28T19:34:21.820Z", "content": "this is temporarily / baseline solution. not the best. (3d cnn should be much better)
\nmore train data /TTA if you use random slice
\n(or random 3d rotation + regular 2d slice)
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F5573c2ac1a06e77722467f2b97722c84%2Fe0a390f43f1888add95e83d9e99c9cb66d403b3e.png?generation=1701200001550844&alt=media\")
def extract_from_image_3d(\n image: torch.Tensor,\n coordinates: torch.Tensor\n) -> torch.Tensor:\n \n\n Parameters\n ----------\n image: torch.Tensor\n `(d, h, )` volume.\n coordinates: torch.Tensor\n `(..., zyx)` array of coordinates at which `image` should sampled.\n Coordinates should ordered zyx, aligned with image dimensions `(d, h, )`.\n Coordinates should array coordinates, spanning `[, -]` \n dimension of length .\n[1] Memory transformers for full context and high-resolution 3D Medical Segmentation
\nhttps://arxiv.org/pdf/2210.05313.pdf
\"Combined, they allow full attention over high resolution images, e.g. 512 x 512 x 256 voxels and above. Experiments on the BCV image segmentation dataset shows better performances than state-of-the-art CNN and transformer baselines, ….\"
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F81a6490b732e022cd6f0678f6c602a51%2FSelection_999(4370).png?generation=1702338190685756&alt=media\")
why you need self-supervised learning
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F16ab1afd3a595132b6488e6dab19dc1e%2FSelection_999(4338).png?generation=1702262797754051&alt=media\")
_50um_LADAF-2020-31_kidney : 1644x1108,2141
", "rawMarkdown": "why you need self-supervised learning\n.png?generation=1702262797754051&alt=media)\n\n_50um_LADAF-2020-31_kidney : 1644x1108,2141\n", "votes": 2, "replies": [ { "id": 2556824, "postDate": "2023-12-11T02:49:32.700Z", "content": "HINT: it is easier to check if this is kidney 5 or 6
", "rawMarkdown": "HINT: it is easier to check if this is kidney 5 or 6", "replies": [ { "id": 2557364, "postDate": "2023-12-11T12:46:43.260Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fb3e354c788cc28eb4ecc8f4c47bfe9cb%2FSelection_999(4356).png?generation=1702298787463603&alt=media\")
\nquite noisy
if the holding cylinder is the same, this gives away the size and voxel resolution of the object
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F8f786a337a36b871e100035520f0e34f%2FSelection_999(4289).png?generation=1702090766447526&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F396b2c54d8a93cd1aab3f9ca168f6468%2FSelection_999(4328).png?generation=1702258596888327&alt=media\")
this is why the vessel are detected as broken
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fc82f66f71452b66518ef361f8c2a58ca%2FSelection_999(4258).png?generation=1701800721703848&alt=media\")
@hengck23 The vessels only look like a hole when viewed by slices, I suppose we actually won't see the faded/broken vessels in the slice (your image on the left), is that right? And the stuff inside, did you find that's for broken vessels only or it's generic?
", "rawMarkdown": "@hengck23 The vessels only look like a hole when viewed by slices, I suppose we actually won't see the faded/broken vessels in the slice (your image on the left), is that right? And the stuff inside, did you find that's for broken vessels only or it's generic? " } ] }, { "id": 2549368, "postDate": "2023-12-05T07:25:52.270Z", "content": "Thank you very much for your sharing, based on your methods and information, here are the experimental data on my side.
\nBy the way, how did you generate the kidney's mask? I haven't used it yet.![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F856355%2F2993476c17278115668affe8cee63e35%2F5101701761042_.pic.jpg?generation=1701761134544839&alt=media\")
\" you generate the kidney's mask?\"
\ni label a few and train a model to label the rest.
you should try the kaggle metric (surface dice) on your validation set as well.
\niou is not good enough to capture cv/lb correlation
try lower threshold as well
", "rawMarkdown": "\" you generate the kidney's mask?\"\ni label a few and train a model to label the rest.\n\n\nyou should try the kaggle metric (surface dice) on your validation set as well.\niou is not good enough to capture cv/lb correlation\n\ntry lower threshold as well" } ] }, { "id": 2529617, "postDate": "2023-11-18T11:39:22.003Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa452e22583e2afc291b2842af7cebc3a%2FSelection_999(3956).png?generation=1700307537301570&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F493812cfdecb2d890397597413ff1cea%2FSelection_999(3957).png?generation=1700307550363042&alt=media\")
what solution to expect
", "rawMarkdown": ".png?generation=1700307537301570&alt=media)\n\n.png?generation=1700307550363042&alt=media)\n\nwhat solution to expect", "votes": 2, "replies": [ { "id": 2530323, "postDate": "2023-11-19T03:10:31.813Z", "content": "very, very important papers and resources:
\n[1] Micro to macro scale analysis of the intact human renal arterial tree with Synchrotron Tomography
\nhttps://www.biorxiv.org/content/10.1101/2023.03.28.534566v1.full.pdf
\nhttps://zenodo.org/records/7708966#.ZGeLVnbMKUl
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Faacfa90372d8c8e608144b066062647d%2FSelection_999(4024).png?generation=1700837683978825&alt=media\")
here is the trick
", "rawMarkdown": ".png?generation=1700837683978825&alt=media)\n\nhere is the trick", "votes": 1, "replies": [ { "id": 2537142, "postDate": "2023-11-24T19:36:07.547Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F0675a83e4fb9dbf880a2f3c2dc38b023%2FSelection_999(4027).png?generation=1700854558385115&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F3dd839bd1bca73bf14f015ceaf74f17f%2FSelection_999(4028).png?generation=1700854641451420&alt=media\")
\nhttps://www.youtube.com/watch?app=desktop&v=DTmVbeJCEBQ
OOPS!!!!
\nunexpected shift in reszing ….
\n![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F4e5901add416e6912b114e17c9296e49%2FPeek%202024-02-03%2014-37.gif?generation=1706942464276991&alt=media\")
can you catch the bug?
\nit seems to be \"easy code\" to change private resolution to public one and then convert the prediction to private one again. make sure you do it correctly!
\n = \n base = \n s = base/\n\n np..rand()<p:\n H,W = .shape\n = cv2.(,dsize=None, fx=s,fy=s)\n = cv2.(,dsize=(W,H))\n![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F8f41cc01426bbb996c6d5b4b7166a99f%2FPeek%202024-02-03%2014-38.gif?generation=1706942949714394&alt=media\")
",
"rawMarkdown": "after bug correction\n"
},
{
"id": 2633884,
"postDate": "2024-02-03T10:48:28.003Z",
"content": "What was the bug here?
", "rawMarkdown": "What was the bug here?" }, { "id": 2634157, "postDate": "2024-02-03T14:45:32.370Z", "content": "How to correct the bug? The public LB score of using resize twice in inference is reduced.
", "rawMarkdown": "How to correct the bug? The public LB score of using resize twice in inference is reduced.", "votes": 1 } ] } ] }, { "id": 2593551, "postDate": "2024-01-09T09:32:03.907Z", "content": "45 degree TTA
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fcbda874375758ee05f1cd451ef85a8cd%2FSelection_999(4539).png?generation=1704792722382609&alt=media\")
Thank you for sharing. I want to know how much the lb score has improved using this 45 degree TTA.
", "rawMarkdown": "Thank you for sharing. I want to know how much the lb score has improved using this 45 degree TTA.", "votes": 1 }, { "id": 2597902, "postDate": "2024-01-12T02:24:08.660Z", "content": "updated
\n make_digonal_slice(shape=[,,]):#D, H, W\n , H, W = shape\n\n = max(H,W)\n = min(H,W)\n = int(.*((L**+L**)**.))\n\n ,y = np.arange(S), np.arange(S)\n\n \n \n\n =[]\n =[]\n h in range(H):\n = x\n = y+h\n = (yy>=) & (yy<H) & (xx>=) & (xx<W)\n valid.sum()>L:\n , xx1 =yy[valid], xx[valid]\n .append([yy1, xx1])\n\n = yy1\n = W--xx1#np.ascontiguousarray(xx1[::-])\n .append([yy2, xx2])\n\n w in range(,W):\n = x+w\n = y\n = (yy>=) & (yy<H) & (xx>=) & (xx<W)\n valid.sum()>L:\n , xx1 =yy[valid], xx[valid]\n .append([yy1, xx1])\n\n = yy1\n = W--xx1\n .append([yy2, xx2])\n\n = coord1+coord2\n coord\n\n\n = make_digonal_slice(shape=[D, H, W])\n\n\n y,x in coord:\n = volume[:,y,x]\n ('slice', slice,resize=.)\n .waitKey()\n\n\n = np.zeros((H, W))\n y,x in coord:\n [y,x]+=\n ('check', check,min=,max=,resize=.)\n .waitKey()\n.waitKey()\njust an initial idea to handle both public and private data. Not tested yet and not sure if it would actually work ????
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F6f061606de1cb79ad24a549ccd48af7c%2FSelection_999(4531).png?generation=1704540731627867&alt=media\")
of course a simpler idea is just upscale private data to public resolution at input, follwed by downscale at output
", "rawMarkdown": "of course a simpler idea is just upscale private data to public resolution at input, follwed by downscale at output", "votes": 2 }, { "id": 2617103, "postDate": "2024-01-24T04:28:23.467Z", "content": "How would you validate this ? Validating on the bottom model would mean validating on resized labels (not optimal) but that's the model that would end up being used on the test data right ?
", "rawMarkdown": "How would you validate this ? Validating on the bottom model would mean validating on resized labels (not optimal) but that's the model that would end up being used on the test data right ? " } ] }, { "id": 2561730, "postDate": "2023-12-14T18:46:30.470Z", "content": "an interesing work: Artery-Vein Separation
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F9e292f946042db3cbad2ff224193d813%2FSelection_999(4407).png?generation=1702579578310237&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F27b9b2806e1ff3034714c2f434c08037%2FSelection_999(4408).png?generation=1702579588732159&alt=media\")
it will be interesting if you :
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fc3dc377198edb933e26b6617570ddf87%2FSelection_999(4424).png?generation=1702675246382938&alt=media\")
https://github.com/antonioguj/bronchinet
\nAirway segmentation from chest CTs using deep Convolutional Neural Networks
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fd498a7572abcc58b241a6b738b1f7aa0%2FSelection_999(4399).png?generation=1702527914761421&alt=media\")
Hip-CT lung vasculature and microstructure youtube presentation by Dr Claire Walsh:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F9cbb48254dd3274c7bcce5d6021aeff1%2FSelection_999(4398).png?generation=1702527034572289&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F7b75a9469be2d16c01efc56f64e77f81%2FSelection_999(4396).png?generation=1702527079702177&alt=media\")
\nhttps://www.youtube.com/watch?v=ehDRFzvwGNY
\nhttps://profiles.ucl.ac.uk/59686-joseph-jacob/publications?favouritesFirst=true&perPage=25&sort=dateDesc&startFrom=25
it would be good if unlabelled data could be used!
", "rawMarkdown": "Hip-CT lung vasculature and microstructure youtube presentation by Dr Claire Walsh:\n\n.png?generation=1702527034572289&alt=media)\n\n .png?generation=1702527079702177&alt=media)\nhttps://www.youtube.com/watch?v=ehDRFzvwGNY\nhttps://profiles.ucl.ac.uk/59686-joseph-jacob/publications?favouritesFirst=true&perPage=25&sort=dateDesc&startFrom=25\n\nit would be good if unlabelled data could be used!" }, { "id": 2554208, "postDate": "2023-12-08T22:40:48.663Z", "content": "there is one method that one can try: supervoxel (3d superpixel) + deep net, since the ground truth is build with 3d floodfill
\nkeywords:Mask2Former, kMaX-DeepLab,: Differentiable SLIC, superpixel-transformer
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F660e76a4b215965a75228c7875066679%2FSelection_999(4284).png?generation=1702075080554884&alt=media\")
train log
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa93a20f8f7fa582a2ff6ed97d700ec43%2FSelection_999(4279).png?generation=1702021715695993&alt=media\")
Hello, I apologize for interrupting you. I have a question about your multiview approach.
\nHow do you ensure that the model can simultaneously see and learn features from different perspectives during multiview training?
\nDo you randomly select different views during data loading for training, or do you have a three-branch network where each branch processes one view and then merges at a certain layer? Or, do you feed all the data from the xy, zy, zx views into the model in each epoch?
", "rawMarkdown": "Hello, I apologize for interrupting you. I have a question about your multiview approach.\n\nHow do you ensure that the model can simultaneously see and learn features from different perspectives during multiview training?\n\nDo you randomly select different views during data loading for training, or do you have a three-branch network where each branch processes one view and then merges at a certain layer? Or, do you feed all the data from the xy, zy, zx views into the model in each epoch?", "votes": 1, "replies": [ { "id": 2561148, "postDate": "2023-12-14T10:05:23.090Z", "content": "random slection will do. i use one unet for all 3 views.
\nnum of slices samples =h+w+d
my next plan:
\n2d axial unet to 3d axial transfomer
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F398b1610236d7dc040c9984face43945%2FSelection_999(4272).png?generation=1701973596052591&alt=media\")
i wonder does it make sense to download other organs and treat them as negative images?
", "rawMarkdown": "i wonder does it make sense to download other organs and treat them as negative images?" }, { "id": 2550388, "postDate": "2023-12-06T02:26:48.760Z", "content": "@hengck23 I am a bit confused how many 2D slices make a 3D volume?
", "rawMarkdown": "@hengck23 I am a bit confused how many 2D slices make a 3D volume?" }, { "id": 2547019, "postDate": "2023-12-03T05:19:37.347Z", "content": "new BM18 results:
\nThe HOAHub uses the High-Throughput Large Field Phase-Contrast Tomography Beamline BM18.
\nhttps://mecheng.ucl.ac.uk/HOAHub/
@hengck23 I trained on kidney_1_dense (original resolution) 5 fold cv and tested on kidney_3_dense (original resolution)using the competition metric …
\nmy CV was 0.82
\nbut my LB is 0.002.
\nDo you have any idea what might be the problem?
@arunodhayan did you get why this was happening
", "rawMarkdown": "@arunodhayan did you get why this was happening", "replies": [ { "id": 2549323, "postDate": "2023-12-05T06:23:12.280Z", "content": "yes, there was a bug in my script during inference
", "rawMarkdown": "yes, there was a bug in my script during inference", "votes": 1 } ] } ] }, { "id": 2542408, "postDate": "2023-11-29T08:54:22.667Z", "content": "@hengck23 do you use binary image to train or convert it to RGB for training?
", "rawMarkdown": "@hengck23 do you use binary image to train or convert it to RGB for training?\n" }, { "id": 2540680, "postDate": "2023-11-27T22:19:58.893Z", "content": "@hengck23 What is the metric you use for validation?
", "rawMarkdown": "@hengck23 What is the metric you use for validation?", "replies": [ { "id": 2540763, "postDate": "2023-11-28T01:24:09.710Z", "content": "version 22
\nhttps://www.kaggle.com/code/metric/surface-dice-metric/notebook
I trained a seresnext26 (I assume that is the model you use), with kidney_1_dense, on original image size (except that I had to pad it to the nearest 32 divisible number), I used no augmentations, I did not tune the learning rate, started at 1e-4, cosine-annealing scheduler, trained for 30 epochs, dice loss from smp, a very basic baseline, saw it overfit and diverge, took the last epoch and calculated the metric on kidney_3_dense
\nTo calculate the metric, I took the predictions on the original size of kidney_3_dense, removed the padding, and converted them into RLE, then I ran it on competition metric
\n![\"\"](\"https://pasteboard.co/YfLE4a0qsHew.png\")
That is the lowest I have seen this metric, I normally get 0.94-0.95 for my finetuned models which scores 0.76+ LB, so I don't know how you have it lower than 0.85 and still that good of a leaderboard, are we calculating the same thing? I just took the notebook you gave, copied the code and pasted it in a file with no change, am I doing some mistake?
", "rawMarkdown": "@hengck23 \n\nI trained a seresnext26 (I assume that is the model you use), with kidney_1_dense, on original image size (except that I had to pad it to the nearest 32 divisible number), I used no augmentations, I did not tune the learning rate, started at 1e-4, cosine-annealing scheduler, trained for 30 epochs, dice loss from smp, a very basic baseline, saw it overfit and diverge, took the last epoch and calculated the metric on kidney_3_dense\n\nTo calculate the metric, I took the predictions on the original size of kidney_3_dense, removed the padding, and converted them into RLE, then I ran it on competition metric\n\n\n\nThat is the lowest I have seen this metric, I normally get 0.94-0.95 for my finetuned models which scores 0.76+ LB, so I don't know how you have it lower than 0.85 and still that good of a leaderboard, are we calculating the same thing? I just took the notebook you gave, copied the code and pasted it in a file with no change, am I doing some mistake?" }, { "id": 2540798, "postDate": "2023-11-28T02:46:06.533Z", "content": "\"0.94-0.95 for my finetuned models \" seems unusually high for me.
\ni will put my validation code and results on public notebook soon for you to check.
\nplease wait for a while.
you probably don't have to mind so much.
\nI know CV vs LB can be different for different kaggler.
\nit is probably due to the different probability threshold (e.g. 0.5,) and threshold used in removal of false positives fp (e.g. contour object of small 2d area or 3d volume)
one may have better CV score locally because less local fp.
\nThe LB may be better or worst becuase of \"sudden\" absence/occurence of fp in hidden test data.
(at such, i think results may not be stable?)
", "rawMarkdown": "\"0.94-0.95 for my finetuned models \" seems unusually high for me.\ni will put my validation code and results on public notebook soon for you to check.\nplease wait for a while.\n\n---\n\nyou probably don't have to mind so much.\n\n---\nI know CV vs LB can be different for different kaggler.\nit is probably due to the different probability threshold (e.g. 0.5,) and threshold used in removal of false positives fp (e.g. contour object of small 2d area or 3d volume)\n\none may have better CV score locally because less local fp.\nThe LB may be better or worst becuase of \"sudden\" absence/occurence of fp in hidden test data.\n\n(at such, i think results may not be stable?)\n\n" }, { "id": 2540828, "postDate": "2023-11-28T03:33:50.043Z", "content": "do you pad or resize the original image to nearest?
", "rawMarkdown": "do you pad or resize the original image to nearest?\n" }, { "id": 2540843, "postDate": "2023-11-28T03:43:20.897Z", "content": "Zeros Padding
", "rawMarkdown": "Zeros Padding" }, { "id": 2540878, "postDate": "2023-11-28T04:26:24.610Z", "content": "\nhttps://www.kaggle.com/code/hengck23/check-kaggle-metric-for-cv
\nexample data and notebook (all public) for
hr min\nkidney_3_dense\n \n \n(LB)\n@harshitsheoran did you find out the reason for your higher CV score?
", "rawMarkdown": "@harshitsheoran did you find out the reason for your higher CV score?" }, { "id": 2543624, "postDate": "2023-11-30T08:37:26.173Z", "content": "Yes, I had height and width switched because I was using PIL the first time instead of cv2 😅, my CV and LB does not completely correlate but I have CV of 0.821 for my 0.782 LB
", "rawMarkdown": "Yes, I had height and width switched because I was using PIL the first time instead of cv2 😅, my CV and LB does not completely correlate but I have CV of 0.821 for my 0.782 LB", "votes": 2 }, { "id": 2543677, "postDate": "2023-11-30T09:15:14.090Z", "content": "\nHow do you 'remove the padding' of kidney_3_dense when scoring?
\nTrain: kidney_1_dense (1303, 912) --> pad zero --> (1536x1536)
\nScore: kidney_3_dense (1706, 1510) --> removed the padding? --> (1536x1536)
I padded (1303, 912) -> (1312, 928) [nearest multiple of 32]
\nI padded (1706, 1510) -> (1728, 1536) [nearest multiple of 32]
\nI always pad zeros on the right and bottom of the image, so removing the padding on predicted mask was as simple as
\nmask = mask[:1706, :1510]
@harshitsheoran you trained on 1303x912 and test on different image size?
", "rawMarkdown": "@harshitsheoran you trained on 1303x912 and test on different image size?" }, { "id": 2543761, "postDate": "2023-11-30T10:41:04.320Z", "content": "Yes, this is what I thought \"original\" resolution meant
", "rawMarkdown": "Yes, this is what I thought \"original\" resolution meant" }, { "id": 2544833, "postDate": "2023-12-01T05:36:58.317Z", "content": "I'm using the same model trained on kidney_1_dense, original image size and using diceloss as model selection because the competition metric is slow. However, I usually get a cv ~0.58 on kidney_3_dense for the best weights. Is this normal? It seems that everyone using this model claims a cv higher than 0.8 is achievable.
", "rawMarkdown": "I'm using the same model trained on kidney_1_dense, original image size and using diceloss as model selection because the competition metric is slow. However, I usually get a cv ~0.58 on kidney_3_dense for the best weights. Is this normal? It seems that everyone using this model claims a cv higher than 0.8 is achievable. \n" }, { "id": 2553035, "postDate": "2023-12-08T01:13:48.127Z", "content": "Hello, I would like to know if there are any methods to prevent scoring errors when using a lower threshold, such as 0.005, apart from using functions like remove_small_objects.
", "rawMarkdown": "Hello, I would like to know if there are any methods to prevent scoring errors when using a lower threshold, such as 0.005, apart from using functions like remove_small_objects.\n\n\n\n\n\n\n" }, { "id": 2553040, "postDate": "2023-12-08T01:21:11.780Z", "content": "in the final solution, your model needs to work well for higher threshold (when your LB improves, you will find that your threshold also increases, i.e. your model is more confident).
\nduring development, you can think of way to remove fp e.g. if you detect the kidney object, you can mask off vessel prediction outside the kidney region.
\nuse local validation to find the best threshold. you should train your model so that the fp is not too high.
", "rawMarkdown": "in the final solution, your model needs to work well for higher threshold (when your LB improves, you will find that your threshold also increases, i.e. your model is more confident).\n\nduring development, you can think of way to remove fp e.g. if you detect the kidney object, you can mask off vessel prediction outside the kidney region.\n\nuse local validation to find the best threshold. you should train your model so that the fp is not too high.", "votes": 3 }, { "id": 2553057, "postDate": "2023-12-08T01:46:38.500Z", "content": "This is very helpful, thank you.
", "rawMarkdown": "This is very helpful, thank you." } ] } ] }, { "id": 2533956, "postDate": "2023-11-22T10:10:48.757Z", "content": "Tips:
\nhere is prediction results for prob<0.25.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fd72f62d167530e54acfc4cf68dbb08db%2FSelection_999(4019).png?generation=1700647828309499&alt=media\")
image resolution used during training?
", "rawMarkdown": "image resolution used during training?", "replies": [ { "id": 2534073, "postDate": "2023-11-22T11:31:32.020Z", "content": "full resolution
", "rawMarkdown": "full resolution", "votes": 1 }, { "id": 2534080, "postDate": "2023-11-22T11:40:04.460Z", "content": "do u mean original image resolution?
", "rawMarkdown": "do u mean original image resolution?" }, { "id": 2534096, "postDate": "2023-11-22T12:00:26.403Z", "content": "yes, for both train and test
", "rawMarkdown": "yes, for both train and test", "votes": 2 }, { "id": 2534101, "postDate": "2023-11-22T12:06:23.303Z", "content": "for inference (submission) you use train resolution?
", "rawMarkdown": "for inference (submission) you use train resolution?" }, { "id": 2534137, "postDate": "2023-11-22T12:53:51.503Z", "content": "currently, i just use the image resolution (i.e. no rescale)
\nbut, it is stated that
\n\"These scans may or may not use a different beamline or resolution from the scans used in the training set.\"
\nhence you should probe the best scale and then decide how to augment
i am surprised that probably only 2d unet is sufficient:
\nblue: ground truth
\nred : predicted (threshold at 0.10)
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa21fce79c1efa5172349b15d6a0a6ec9%2FPeek%202023-11-27%2008-04.gif?generation=1701043499019411&alt=media\")
it is pretty obvious that you can use low threshold e.g. 0.10 for early development and get lb score of say 0.6 to 0.7
\nbut for high lb score of 0.8,
\n```
", "rawMarkdown": "i am surprised that probably only 2d unet is sufficient:\n\nblue: ground truth\nred : predicted (threshold at 0.10)\n\n\n\nit is pretty obvious that you can use low threshold e.g. 0.10 for early development and get lb score of say 0.6 to 0.7\n\nbut for high lb score of 0.8, \n- you need high threshold to remove false positive (i.e. you need a better confidence deep net)\n- you need to filter off the false positive with post processing or some second stage 3d filtering deep net.\n\n\n \n\n```", "votes": 2, "replies": [ { "id": 2540779, "postDate": "2023-11-28T02:14:54.900Z", "content": "i previously show the false positive.
\nHere, i show the false negative (miss)
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F126f8735d4a7adc180dfa773f70ed50a%2FSelection_999(4088).png?generation=1701137639848452&alt=media\")
broken vessel seems easy to fix.
\nHence at least for the public LB, i think greater than 0.90 is very likely
Hi, Heng
\nWhat do you use to visualize 3d volumes?
\nThanks!
@sakvaua see code at :
\nhttps://www.kaggle.com/code/hengck23/check-kaggle-metric-for-cv
search for the below:
\n## visualisation to produce results above\n :\n pyvista pv\n pl = pv.Plotter()\n ...\nThanks! I like your animations!
", "rawMarkdown": "Thanks! I like your animations!" }, { "id": 2544960, "postDate": "2023-12-01T07:48:19.077Z", "content": "from the paper:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fd4620fdeacf33576aab05f410e636fe7%2FSelection_999(4172).png?generation=1701416838855852&alt=media\")
In my experiments, this can be solved by data augmentation, non-uniform scale to squash the vessel to make them flat, elongated
", "rawMarkdown": "from the paper:\n\n.png?generation=1701416838855852&alt=media)\n\nIn my experiments, this can be solved by data augmentation, non-uniform scale to squash the vessel to make them flat, elongated", "votes": 2 } ] } ] }, { "id": 2529362, "postDate": "2023-11-18T06:53:51.610Z", "content": "were you able to differentiate between dense and sparse annotation. what do you think will be best to use
", "rawMarkdown": "were you able to differentiate between dense and sparse annotation. what do you think will be best to use", "replies": [ { "id": 2529512, "postDate": "2023-11-18T10:19:24.923Z", "content": "i haven't look into details yet. i think sparse exclude smallest vessels.
\n(to know the difference, i can train on sparse and then use the model to detect dense, and vice versa)
my plan is to use only dense first to build a model.
\nthen think of how use sparse set as partial label set. (e.g. self supervised + label)
", "rawMarkdown": "i haven't look into details yet. i think sparse exclude smallest vessels.\n(to know the difference, i can train on sparse and then use the model to detect dense, and vice versa)\n\nmy plan is to use only dense first to build a model.\n\nthen think of how use sparse set as partial label set. (e.g. self supervised + label)", "votes": 1, "replies": [ { "id": 2529518, "postDate": "2023-11-18T10:29:22.383Z", "content": "blue: sparse
\npink: difference of dense, sparse
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F55331545290675144f2535cef4e943ff%2FSelection_999(3955).png?generation=1700303301787011&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F36d1d4753e056c4c65feb2bb129e8aab%2FSelection_999(3954).png?generation=1700303314539871&alt=media\")
If sparse annotations are used for the small vessel, it may be beneficial to merge dense and sparse annotations, allowing the model to effectively capture both during prediction.
", "rawMarkdown": "If sparse annotations are used for the small vessel, it may be beneficial to merge dense and sparse annotations, allowing the model to effectively capture both during prediction.", "votes": 1 } ] } ] }, { "id": 2600229, "postDate": "2024-01-13T13:23:23.823Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2600227, "postDate": "2024-01-13T13:21:49.707Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2584348, "postDate": "2024-01-02T20:44:54.860Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 2583472, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-02T08:56:51.607000", "content": "i realize something important.
\ni checked the fast surface dice loss and i think we can directly optimize this loss in gradient descent!
i) use 3 channel input and predict 3 channel output
\nii) refer to the fast surface dice code by @junkoda . you can apply same marching cube algorithm (2x2x2) to compute soft surface dice loss
instead of unfold 3d conv with 256 kernels may be slightly faster
\nthis may be important because i realize the reason for missing small vessel is annotation error (even for the 100% dense label, not all smallest vessel are annotated). this causes the following
\nFrom my understanding the surface dice metric requires thresholding of predictions hence we cannot directly optimize it.
\nApplying the metric on probabilities does not seem to yield meaningful results but I did not dig very deeply nor look into the math.
here is the trick:
\n(it works because we are using tolerance =0)
if tolerance is not zero, then we have to modify the softmax loss to KL divergence loss where the ground truth probabiliy = 1 - surface distance
", "votes": 3, "replies": [] }, { "id": 2584045, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-02T16:39:31.320000", "content": "surface dice code
\n
BCE is really a bad loss
\n
some other good boundary/surface loss from other papers:
\n[1] Boundary loss for highly unbalanced segmentation
\nhttps://arxiv.org/pdf/1812.07032.pdf
\nhttps://github.com/LIVIAETS/boundary-loss
[2] Boundary Difference Over Union Loss For Medical Image Segmentation
\nhttps://arxiv.org/pdf/2308.00220.pdf
\nhttps://github.com/sunfan-bvb/BoundaryDoULoss
google new paper!!
\nBoundary Attention: Learning to Find Faint Boundaries at Any Resolution
\nhttps://arxiv.org/pdf/2401.00935.pdf
Thanks for sharing, have you try this loss （your surface dice code）, how effective was it?
", "votes": 0, "replies": [] } ] } ] }, { "id": 2537053, "author_name": "hengck23", "author_url": "", "post_date": "2023-11-24T17:39:09.587000", "content": "kidney mask dataset
\nhttps://www.kaggle.com/datasets/hengck23/blood-vessel-segmentation-kidney-mask
Thanks for putting this together, I was just looking at doing something like this yesterday before getting pulled in to holiday shenanigans. I re-scaled the data back to the full image size, RLE encoded it, and saved it back to the training csv for anyone interested in using it without resizing on load: https://www.kaggle.com/datasets/squidinator/sennet-hoa-kidney-13-dense-full-kidney-masks
", "votes": 2, "replies": [] }, { "id": 2537722, "author_name": "hengck23", "author_url": "", "post_date": "2023-11-25T12:40:51.590000", "content": "this is the code to get lb710+ using kidney mask
\n vessel, kidney = 0,0\n with torch.cuda.amp.autocast:\n with torch.no_grad:\n vm, km = net \n vessel += vm\n kidney += km\n\n TTA here \n\n \n vessel = vessel.float\n kidney = kidney.float\n\n _size = len\n for b in range:\n mk = kidney[b, 0]\n mk = choose_biggest_object \n\n mv = vessel[b, 0]\n mv = )\n \n p = \n rle = rle_encode\n df_data.append\nWhat is TTA?
", "votes": 0, "replies": [] }, { "id": 2538656, "author_name": "coderRKJ", "author_url": "", "post_date": "2023-11-26T10:56:20.430000", "content": "TTA is Test Time Augmentation. Best summary is in this paper: Better Aggregation in Test-Time Augmentation
\n\n\nTest-time augmentation -- the aggregation of predictions across transformed versions of a test input
\n
TL;DR: Make N augmentations of incoming test data and then combine them (e.g. taking average of the N outputs) to output the final results. It reduces variance in model output.
", "votes": 0, "replies": [] } ] }, { "id": 2538673, "author_name": "Kafka Tamura", "author_url": "", "post_date": "2023-11-26T11:26:38.900000", "content": "Can you please explain this a bit? What are these masks for? In the train datasets, don't we already have dense masks for kidney 1 & 3? Please excuse me if this is mundane. I'm just starting out.
", "votes": 1, "replies": [] }, { "id": 2544657, "author_name": "lhwcv", "author_url": "", "post_date": "2023-12-01T02:18:16.403000", "content": "Thanks for Sharing. How you create this kidney mask dataset？
", "votes": 1, "replies": [] } ] }, { "id": 2568510, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-20T15:35:22.913000", "content": "20-dec:
\nrecepie for LB 0.848
\ndetails at the other posts below ….
\ni am surprised that it can get LB 0.862 by training for long iterations (but CV does not reflect the change … maybe because of wrong annotation on kidney2 and the segmentation sparsity (i.e. incomplete label)
\ni.e. there is probably no reliable local validation set in this competition
", "votes": 7, "replies": [ { "id": 2568889, "author_name": "samshipengs", "author_url": "", "post_date": "2023-12-21T00:10:57.317000", "content": "maybe i missed it, but any reason just normal unet2d would be suffice? any thoughts on 3D approach?
", "votes": 0, "replies": [] }, { "id": 2575634, "author_name": "Time Master", "author_url": "", "post_date": "2023-12-27T03:24:57.123000", "content": "Hi, I don't understand what does xy,zx,zy mean，could you please explain it, thanks!
", "votes": 2, "replies": [ { "id": 2576362, "author_name": "Ivan Egorov", "author_url": "", "post_date": "2023-12-27T16:50:12.083000", "content": "As you are dealing with 3d images, you can view them at different sides, so there are XY, ZX and ZY. Think of that as a cube which you can look at from different angles.
", "votes": 3, "replies": [] }, { "id": 2576585, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-27T22:33:40.620000", "content": "
more infor
", "votes": 3, "replies": [] } ] }, { "id": 2582421, "author_name": "Muhammad Haseeb Ahmad", "author_url": "", "post_date": "2024-01-01T15:19:38.600000", "content": "did you try using a combination of these strats, or did you get excellent results trying each of them individually?
", "votes": 0, "replies": [] } ] }, { "id": 2565193, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-17T18:47:04.473000", "content": "correct way to do validation:
\nNOTE:
\n
another tip:
\nmy previous two winning kaggle solution has huge shake up. I could win because i kept to the follwing principles. instead of the best parameter (e.g. threshold) for the solution:
in summary, in addition to best results, think of ways to
\nhttps://www.kaggle.com/code/junkoda/fast-surface-dice-computation
\nbased on the fast surface dice computation by @junkoda, i did an extensive study on the effect of threshold and the local lb metric. For each model after the training epoch, i made measurements.
\nconclusion:
\n
I'm using a combination of 0.5Dice loss and 0.5BCE loss. The loss is continuously decreasing, but it seems that the surface dice and LB are also increasing. Perhaps Dice loss is more suitable?
", "votes": 1, "replies": [] }, { "id": 2556937, "author_name": "Snorf", "author_url": "", "post_date": "2023-12-11T05:02:50.523000", "content": "Is the model trained on xy xz yz or only inferred on xy xz yz?
", "votes": 1, "replies": [ { "id": 2556969, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-11T05:41:15.430000", "content": "model is trained in xy xz yz
", "votes": 4, "replies": [] } ] }, { "id": 2562703, "author_name": "Lingxi", "author_url": "", "post_date": "2023-12-15T16:01:34.850000", "content": "Hello, may I ask how you do multi-view training during the training process, is it cycling three axes training in one epoch? I put the data of the three views into a dataset, set the batchsize to 1, and shuffle it, but it doesn't seem to work well. Could you please make a reply? Thank you very much
", "votes": 2, "replies": [ { "id": 2562785, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-15T16:46:53.713000", "content": " ():\n train_id=[]\n n name:\n d = meta_data[n]\n D,H,W = d.image.shape\n train_id += [ (d.name, i, ) i (H) ]\n train_id += [ (d.name, i, ) i (D) ]\n train_id += [ (d.name, i, ) i (W) ]\n train_id\n\ntrain_id = make_train_id(...)\n\n\n ():\n ():\n self.sample_id = sample_id\n self.augment = augment\n self.length = (self.sample_id)\n\n unique_name=[]\n name,i,axis sample_id:\n name unique_name:\n unique_name.append(name)\n self.unique_name=(unique_name)\n\n ():\n string = \n string += \n string += \n string\n\n ():\n self.length\n\n ():\n name,i,axis = self.sample_id[index]\n\n d = DATA_META[name]\n D,H,W = d.image.shape\n\n axis == :\n image = d.image[i]\n vessel = d.vessel[i]\n axis == :\n image = d.image[:, i]\n vessel = d.vessel[:, i]\n axis == :\n image = d.image[:, :, i]\n vessel = d.vessel[:, :, i]\n\n image = np.ascontiguousarray(image)\n vessel = np.ascontiguousarray(vessel)\n\n self.augment :\n image, vessel = self.augment(image, vessel)\n\n image = np.ascontiguousarray(image)\n vessel = np.ascontiguousarray(vessel)\n \n\n r = {}\n r[] = index\n r[ ] = (name,i,axis)\n r[ ] = torch.from_numpy(image).()\n r[] = torch.from_numpy(vessel).()\n r\nThis seems similar to doing a three-view dataset offline, whether you used a batchsize of 1 or padded the different viewing images to the same size.
", "votes": 1, "replies": [] } ] } ] }, { "id": 2555308, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-09T20:32:34.497000", "content": "example of simple 3d flood fill
\n
iteration in range():\n \n t in range(,,-): \n (t)\n = dense[t+]\n = image[t+]\n = image[t]\n\n = curr_image[:-,:-]\n = prev_mask*prev_image\n\n =\n = \n += np.abs(m-prev[:-,:-])<th #[ , ]\n += np.abs(m-prev[: ,:-])<th #[ , ]\n += np.abs(m-prev[:-,:-])<th #[-, ]\n += np.abs(m-prev[:-,:-])<th #[ ,-]\n += np.abs(m-prev[: ,:-])<th #[ ,-]\n += np.abs(m-prev[:-,: ])<th #[-,-]\n += np.abs(m-prev[:-,: ])<th #[ , ]\n += np.abs(m-prev[: ,: ])<th #[ , ]\n += np.abs(m-prev[:-,: ])<th #[-, ]\n\n = diff>\n \n\n += grow\n \n ('predict',(predict>).astype(np.float32))\n .waitKey()\nIMPORTANT!!!!!
\nThis is the black magic and the trick to winning!!!!
results of 3d flood fill with seed = one pixel
\n
this reminds me of 3d SAM (segment anything model)
\nif so, we can use self-supervised learning to train the encoder
the seed will be prompt.
\nthe decoder is to perform floodfill segmentation
", "votes": 2, "replies": [ { "id": 2544343, "author_name": "hengck23", "author_url": "", "post_date": "2023-11-30T19:01:08.027000", "content": "volumetric rendering of image volume
\n
it shows the uniform intensity of walls of the vessel. that is why flood fill works well
", "votes": 1, "replies": [] }, { "id": 2544387, "author_name": "hengck23", "author_url": "", "post_date": "2023-11-30T19:32:36.910000", "content": "
this solves the puzzle why there are so much noise in prediction.
\nwe need some median filtering layer?
\nhttps://github.com/llmpass/medianDenoise
hyptersis thresholding prediction as an alterantive to flood filling on input
", "votes": 3, "replies": [] }, { "id": 2547016, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-03T05:16:38.150000", "content": "nice rendering
\nhttps://www.youtube.com/watch?v=wZbfNJxTxqc
wrong annotation !!!!!
\n
i think there is offset bug in annotion for kidney2
\n
shift error happens in hidden test datat ground truth before in kaggle competitions…
\nanyone want to probe that?
only top half of kidney 2 annotation are shifted
\nTOP:
\n
BOTTOM:
\n
pl = pv.\n\n mhit = pv.).T).glyph(geom=pv.)\n mfp = pv.).T).glyph(geom=pv.)\n mmiss = pv.).T).glyph(geom=pv.)\n pl.add\n pl.add\n pl.add\n pl.show\nmy estimate at one week before the competition:
\nfinal top-1 private leaderboard: private lb 0.890 to 0.880
\ngold cut-off : private lb 0.875 to 0.880
It will be interesting to see how it plays out. I am looking forward to people's write ups so I can understand how they go from 0.8 -> ~0.86 on the leaderboard. I've struggled to get my 2D models (based on SeResNet) to score higher than 0.8. I have learned a lot, but there is plenty more to learn.
", "votes": 0, "replies": [ { "id": 2623007, "author_name": "chemdatafarmer", "author_url": "", "post_date": "2024-01-27T21:02:10.803000", "content": "Actually, I have struggled to get train and validation dice scores over 0.84 (kidney 1 train, kidney 3 val) without overfitting
", "votes": 0, "replies": [] }, { "id": 2623113, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-28T01:36:22.463000", "content": "\"so I can understand how they go from 0.8 -> ~0.86 on the leaderboard. \"
\n\"I have struggled to get train and validation dice … without overfitting\"
answer to both : train augmentation.
\ngenerally(i.e. this and other competitions) all image train augmentation can be classified as
\nsome make the results very worst, some will improve significantly
\nvisualise, visualise, visualise
\nThank you @hengck23 for always being so helpful on these forums! I noticed that the train augmentation helps improve the consistency between the train, test, and even leaderboard. Perhaps I need to try more! I appreciate the advice to inspect early epochs. I hadn't done that, I typically just save the best weights and inspect the result. I'll give it a try :).
", "votes": 0, "replies": [] }, { "id": 2626979, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-30T12:08:10.753000", "content": "visualisation for the first 0.25 epoch
\n
\n
Very interesting! I've been working on doing similar after your comment. Your advice has been very helpful, I've been implementing it and hoping it can help me move up on the leaderboard before the final :)
", "votes": 0, "replies": [] } ] }, { "id": 2623085, "author_name": "Cody_Null", "author_url": "", "post_date": "2024-01-28T00:27:55.863000", "content": "This one has been a heart breaker, I have been stuck with my current score for about a month now and it seems unless I come up with something I will be just shy of gold. I am interested to see what others did because I feel like I have tried so many things that just don’t seem to help.
", "votes": 2, "replies": [ { "id": 2623137, "author_name": "chemdatafarmer", "author_url": "", "post_date": "2024-01-28T02:15:48.037000", "content": "I understand. I got really excited about a custom loss function I found in the literature and for my models below 0.8 on the public leaderboard it always seemed to improve the score, but when I tried it with the model that scored well it didn't improve. I keep hitting a wall. Still, it's been a fun competition! I've never worked on a proper segmentation problem before and I've learned a lot! Good luck getting into gold as the leaderboard shifts with the private data.
", "votes": 1, "replies": [] }, { "id": 2623175, "author_name": "Cody_Null", "author_url": "", "post_date": "2024-01-28T03:23:09.900000", "content": "Thank you! Good luck to you too! Yes the learning is what is important and I have definitely done a lot of that! Good to get reminders every once in awhile!
", "votes": 1, "replies": [] } ] }, { "id": 2625364, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-29T11:09:57.290000", "content": "one more information. from the paper: if train = kidney1 upper, valid = kidney1 lower, surface dice is 0.95.
\nBut i think that is not using TTA.
\n(i haven't done experiment on upper/lower split).
I estimate with TTA and better processing, surface dice is 0.96+.
\nThis means that under perfect data (i.e. no domain shift), expect a 0.03+
", "votes": 0, "replies": [ { "id": 2625675, "author_name": "Muhammad Haseeb Ahmad", "author_url": "", "post_date": "2024-01-29T14:38:28.967000", "content": "I noticed something today which you might find interesting.
\nFor my validation set, which is kidney 3 dense, I generated a submission.csv using all the images.
\nThen I broke it into 10 samples to calculate the surface dice score, each sample with a step size of 10, such that my first sample would consider img 0, 10, 20 etc., the second would consider img 1, 11, 21 and so on.
\nEach individual sample scored above 0.94, and the avg was over .94 as well. However, when I calculated dice score over the full CV set without breaking it into sub samples, the dice score became 0.86. This has been confusing me. Im using @Jun Koda's method. https://www.kaggle.com/datasets/junkoda/sennet-score.
\nObviously my actual score on the public test set is around .835, which is much closer to .86 than .94.
\nI thought that the surface dice scores are being averaged over the sample, but that doesn't seem to be the case.
\nDo you have any thoughts?
What you are doing does not make sense -> the competition metric is computed over consecutive frames: the dice score is computed on the surface area defined by two consecutive frames. So if you compute the score with images separated by 10 images this will create a very poor ground truth for the surface area.
\nIf you wish to see you scores on subsets, you should take adjacent frames like 0 to 100, then 100 to 200 etc… The local scores should then average to the global score.
", "votes": 2, "replies": [] }, { "id": 2625774, "author_name": "Muhammad Haseeb Ahmad", "author_url": "", "post_date": "2024-01-29T15:26:32.107000", "content": "Wow, that clarifies so much for me. Thank you very much. I suppose dice loss is very different.
\nI have learned a valuable lesson.
i wonder did anyone try this:
\nThat's a really good idea, will try and let everyone know the results here
", "votes": 3, "replies": [] }, { "id": 2626640, "author_name": "HB", "author_url": "", "post_date": "2024-01-30T06:33:05.297000", "content": "from the paper: if train = kidney1 upper, valid = kidney1 lower, surface dice is 0.95.
\nCan you tell me about this paper?
\nRelabeling \"kidney3 sparse\" will help, but I don't think it's easy to get a perfect model by using k1 and k2 for validation in \"kidney3 dense\" training. I'm very curious about how much validation data was used for training in the above paper, including the specific training model.
", "votes": 0, "replies": [] } ] } ] }, { "id": 2589018, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-06T01:16:46.003000", "content": "wrong strategy?
\ni have been trying detect small vessels improve LB score becuase stated paper this one problem.\nBut i realize this may be a good strategy?\n\ni use train= kidney1 dense + kidney3 dense\nvalidation= kidney2 sparse ( slice >= avoid annotation shift )\n\ni note recent popular public kernel based seresnext50_32x4d dunet which can easily object lb . seresnext50_32x4d has imagenet top1 range ~. My experiments shows seresnext50_32x4d dunet has low hit rate (>), has very low fp rate (~) well. i am surprise can lb single model.\n\ni have been using nextVIT base transformer, imagenet top1 ~, lb . It has very high hitrate (>) low fp rate (~).\n\n turns out lb based surface pixels, small vessel has less surface ( contribute little lb score) !!!\ne.g. kidney3 sparse (% annotation) has surface dice compared kidney3 dense\n\n other , wrong choice metric host interest detect small vessel\nupdate on 06-JAN
\ntrick to get lb 0.87+
make a base model in the range of 0.864 (e.g. either seresnext50_32x4d or nextVIT base 2d unet)
ensemble of the two will give 0.87+
ensemble improve lb score:
Thank you for sharing. Different models might have different thresholds. How do you perform model fusion? Do you average the probabilities and then use a compromise threshold?
", "votes": 2, "replies": [] }, { "id": 2589337, "author_name": "wpl", "author_url": "", "post_date": "2024-01-06T09:44:17.340000", "content": "Thank you for sharing. I would like to know approximately how long it takes for a single model of amp inference, and would it be faster to use TensorRT instead
", "votes": 1, "replies": [] }, { "id": 2589420, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-06T11:20:49.200000", "content": "seresnext50_32x4d : 2hr +
\nnextVIT base 2d : 3hr +
i haven't use tensorRT yet
", "votes": 1, "replies": [] }, { "id": 2589498, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-06T12:40:51.307000", "content": "
i have not decided. but the threshold is smiliar for same validation set even for different model
", "votes": 4, "replies": [] }, { "id": 2589696, "author_name": "wpl", "author_url": "", "post_date": "2024-01-06T15:17:07.227000", "content": "Thank you for answer.Is this the duration of reasoning for only one perspective without TTA? When I take TTA from three perspectives, seresnext50_ 32x4d took approximately 6 hours to reason
", "votes": 1, "replies": [] }, { "id": 2589706, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-06T15:25:31.097000", "content": "time is for infer with the usual xy,zx,zy + 5 TTA
", "votes": 1, "replies": [] } ] }, { "id": 2590673, "author_name": "Theo Viel", "author_url": "", "post_date": "2024-01-07T11:19:37.070000", "content": "@hengck23 How do you compute your FPR ? Your definition probably differs from the one I'm used to regarding normalization.
\nDo you use (pred * (1 - truth)).sum() / truth.sum() ?
\nSame for TPR, you are using (pred * truth).sum() / truth.sum(), right ?
Thanks!
", "votes": 1, "replies": [ { "id": 2590803, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-07T12:53:54.400000", "content": "def np_metric(, truth):\n p = (>0.5)\n t = (truth>0.5)\n hit = (p*t).()\n fp = (p*(1-t)).()\n t_sum = t.()\n p_sum = p.()\n hit, fp, t_sum, p_sum\n\n#-------------------------------------\n = >0.5\nhit, fp, t_sum, p_sum = np_metric(, truth)\nhit = hit/t_sum\nfp = fp/p_sum\nprecision and recall are better words
", "votes": 4, "replies": [] }, { "id": 2590930, "author_name": "Theo Viel", "author_url": "", "post_date": "2024-01-07T14:46:00.647000", "content": "Gotcha, thanks.
\nFYI, those are my scores on kidney_2:
\n\nhit: 0.9549\nfp: 0.0507\nCV: 0.84\nLB ~0.80\n
\nMy LB scores are lagging behind compared to yours but my CV scores are close. Not sure if it is because I have too many FPs or not.
submit with different thresholds.
\nCV and LB correlations are weak. but i do see a general linear relationship: lb = cv + bias
\n(cv and lb don't share the best optimal threshold, which is a** problem for us to avoid shakeup**)
\"FYI, those are my scores on kidney_2:\"
\nThere are annotation errors in kidney 2. so your lb score should be better if your threshold is correct
", "votes": 2, "replies": [] }, { "id": 2591184, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-07T17:50:56.710000", "content": "@theoviel
\ni think your metric are for model without TTA.
you need to do TTA (it is common that 2d Unet can completely miss vessel in one view and detect in another). TTA may not improve CV, but it should improve LB.
", "votes": 3, "replies": [] } ] } ] }, { "id": 2554978, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-09T15:30:08.363000", "content": "trick to get lb 0.835
\n
def forward(self, batch):\n x = batch['image']\n x = x.(-1,3,-1,-1) \n B, C, , W = x.shape\n\n = []\n xx = self.stem0(x); .(xx)\n xx = F.avg_pool2d(xx,kernel_size=2,stride=2)\n xx = self.stem1(xx); .(xx)\n\n = self.encoder #convnext\n x = .(x);\n\n x = .stages[0](x); .(x)\n x = .stages[1](x); .(x)\n x = .stages[2](x); .(x)\n x = .stages[3](x); .(x)\n ##[(f'encode_{i}', .shape) i, enumerate()]\n\n last, = self.decoder(\n feature=[-1], skip=[:-1][::-1]\n )\n ##[(f'decode_{i}', .shape) i, enumerate()]\n ##('last', last.shape)\n\n vessel = self.vessel(last)\nyou can go to super-resolution
\n
HINT:
\nalternatively, you can train a real super-resolution net (you have 20um imags from the HIP-CT website)
local CV/LB
\n
lb0.829 …
\nnormalisation is the key
\n norm_by_percentile(volume, low=, high=., alpha=.):\n = np.percentile(volume,low)\n = np.percentile(volume,high)\n = (volume-xmin)/(xmax-xmin)\n :\n [x>]=(x[x>]-)*alpha +\n [x<]=(x[x<])*alpha\n \n x\nnormalised by volume/subvolume, not image
", "votes": 6, "replies": [ { "id": 2554394, "author_name": "Arunodhayan", "author_url": "", "post_date": "2023-12-09T05:50:07.933000", "content": "@hengck23 - is this the right way to read 3D volume because i get negative stride error
\n (torch.utils.data.Dataset):\n ():\n self.img_voxel = create_3d_voxel(img_voxel_paths) \n self.msk_voxel = create_3d_voxel_msk(msk_voxel_paths) msk_voxel_paths \n self.transforms = transforms\n\n ():\n self.img_voxel.shape[] \n\n ():\n img = self.img_voxel[index, :, :].copy() \n img = np.expand_dims(img, axis=) \n\n self.msk_voxel :\n msk = self.msk_voxel[index, :, :].copy() \n msk = np.expand_dims(msk, axis=) \n\n self.transforms:\n data = self.transforms(image=img, mask=msk)\n img = data[]\n msk = data[]\n\n torch.tensor(img, dtype=torch.float32), torch.tensor(msk, dtype=torch.float32)\n :\n self.transforms:\n data = self.transforms(image=img)\n img = data[]\n\n torch.tensor(img, dtype=torch.float32)\ndef do_random_flip_rotate(, vessel):\n ..rand()<:\n = .flip(, axis=) #horizontal\n vessel = .flip(vessel,axis=)\n ..rand()<:\n = .flip(, axis=)\n vessel = .flip(vessel,axis=)\n ..rand()<:\n k = ..choice([,,])\n = .rot90(, k, =[,])\n vessel = .rot90(vessel,k, =[,])\n\n = .ascontiguousarray()\n vessel = .ascontiguousarray(vessel)\n , vessel\nstill i face the same error
", "votes": 0, "replies": [] } ] } ] }, { "id": 2541221, "author_name": "hengck23", "author_url": "", "post_date": "2023-11-28T09:58:56.980000", "content": "Yellow: MISS
\n
\nGreen: FP
\n
\nRed: HIT
\n
\nALL
\n
beware !!!
\nit is not any vessel, but those from \" arterial vascular tree \"
data page: \"kidney_1_dense - The whole of a right kidney at 50um resolution. The entire 3D arterial vascular tree … \"
\nhere you can see that not all vessel are annotated (kidney dense 1)
\n
i was wondering do we need to separate arterial and venous?
\n
i am surprised that the deep CNN nework can somhow differentate between venous and arterial (target) vessel from one imgae.
\ncan anyone explain why?
\n(e.g. they have different apperance due to dye/contrast when the Hip CT scan is taken ??? or arterial vessel have thicker wall ???)
The structure of the membrane differs significantly between arteries and veins. Arteries are richer in elastic fibres and thicker than veins. The walls of arteries and veins are very different, as can be seen in the linked histological image (HE-stain).
\nhttps://www.kidneypathology.com/English_version/Vessels_histology.htm
@yosukeyama
\n\"The walls of arteries and veins are very different, \"
Thanks.
\nThat is good news
\nso it seen that we need not grow the tree from scratch. just pure detection (+ post processing) is probably enough.
very bad news !!!
\ni try to do some probing.
public lb set: kidney 5
\nprivate lb set: kidney 6
you may want to verify yourself ( just in case my code is buggy)
", "votes": 6, "replies": [ { "id": 2537624, "author_name": "YumeNeko", "author_url": "", "post_date": "2023-11-25T11:05:56.733000", "content": "I think your guess is correct because I had the same result when I did LB Probing before. (Public is calculated on kidney_5 only, Private is calculated on kidney_6 only.)
\nI think it was reasonable to separate Public and Private by kidney_5 and kidney_6 because it is difficult to calculate metrics with such a high computational load more times than necessary.
I agree with you that the information on the test dataset is rarely publicly available, which causes Big Shake in the worst case scenario.
", "votes": 3, "replies": [ { "id": 2537870, "author_name": "YYama", "author_url": "", "post_date": "2023-11-25T15:06:39.347000", "content": "It was not hard to imagine that this kind of data split is being used.
\nIf the resolution is different, I believe it essentially becomes a completely different image. Since we only have two types of resolution available, appropriate validation is very difficult. While it's not entirely unfeasible, there is a concern that it will turn into a competition heavily influenced by luck. In this case, the top solutions may not necessarily be the superior solutions…
", "votes": 1, "replies": [] }, { "id": 2547015, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-03T05:10:06.350000", "content": "what if private kidney6 (500 images) is actually like kidney1 voi set at about 5.2um resolution.
\ni cannot think of a reason why kidney1 voi set is included in the train set ….
Something funny is my best score 0.62 is from mixing everything… I will probably remove the sparse and see if that makes a difference
", "votes": 0, "replies": [] } ] } ] }, { "id": 2612593, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-21T13:42:56.723000", "content": "experiments on scale … yet another proof that it is FP that matters …
\n
OOPS!!! there is a metric bug
\n(this is behaviour for kidney2. maybe different behaviour for different kidney)
Hello @hengck23,
\nSince your conclusion is that fp matters, have you considered using the TverskyLoss ?
With it you can weight the impact of FP and FN, making the loss closer to the metric. What do you think ?
", "votes": 1, "replies": [] } ] } ] }, { "id": 2589059, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-06T03:02:02.963000", "content": "an interesting idea
\n
How long does it take for you to inference with Transformer, it seems that P100 without fp16 inference is very slow for Transformer, do you have any suggestions for this?
", "votes": 3, "replies": [ { "id": 2568996, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-21T02:45:12.713000", "content": "you can google for fast transformer. i use nextvit, which was used in kaggle rnsa breast mamnographs before. it takes 3to4 hr and can be speedup further by tensort rt to 3 hr
", "votes": 2, "replies": [ { "id": 2569002, "author_name": "lhwcv", "author_url": "", "post_date": "2023-12-21T02:53:01.587000", "content": "Thank you so much!
", "votes": 0, "replies": [] }, { "id": 2569013, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-21T03:04:22.723000", "content": "update: nextVIT-base can get LB0.682 3hr 20 min (just complete in few minutes along so i can get the timing in minutes).
\nit uses 5xTTA + 3 axis (xy,yz,xz) in P100
", "votes": 2, "replies": [] }, { "id": 2593750, "author_name": "Rafael Javier Cruz", "author_url": "", "post_date": "2024-01-09T12:30:04.533000", "content": "Seems that nextVIT is a kind of detection model. Should I modify the configuration of Q1&Q2 Net for segmentation task when using? Thanks!
", "votes": 0, "replies": [] } ] }, { "id": 2569004, "author_name": "slime", "author_url": "", "post_date": "2023-12-21T02:55:23.350000", "content": "Try fp16 + 2xT4 + nn.DataParallel, for me it's faster than using single P100
\nThe model I choose is very slow [maxvit-small],
\nand scoring single checkpoint with xy/yz/xz + 5tta takes 8 hours…
should try Next-ViT I think :D
", "votes": 0, "replies": [ { "id": 2569034, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-21T03:23:30.103000", "content": "
the fastest i can find
", "votes": 2, "replies": [] } ] } ] }, { "id": 2583079, "author_name": "hengck23", "author_url": "", "post_date": "2024-01-02T03:59:54.163000", "content": "did we just find the missing labels (for sparsely annotated ground truth labels)?
\n
GJ. I believe both sparse and dense have labeling issues (imho, cuz I`m not a doctor). Looked through and found even pretty big vessels missed. Thx for sharing.
\n
yes you are correct.
\nan easy way to prove label error is to check train error after overfitting.
\nnow if the train error cannot get to zero, waht can be the reason?
\nExactly. Heavy backbones perform nearly identically as small ones
", "votes": 0, "replies": [] } ] } ] }, { "id": 2560936, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-14T06:01:20.913000", "content": "just an idea:
\ntakes 2 nearby slices and generate vessels in between, both image and mask
\n
i hope there is no bug, but mixing kidney 1 and 3 in train make LB (much worse)
\nreference: train with kidney 1 only: LB 0.757
submit inference mode: add TTA 2xflip,3xrot90 only
\n
since there is no validation set, i just train to the same number of iterations from previous experiments.
\nthis may not be optimal (and hence not optimal lb)???
I trained initially on kidney 1, by adding also kidney 3 improved my LB from 0.65 -> 0.7. I do not know how to compare their resolutions but looks like they have pretty similar ones with 50um and 50.16um. The question would be if LB is also 50um and private something completly different
", "votes": 2, "replies": [ { "id": 2542830, "author_name": "hengck23", "author_url": "", "post_date": "2023-11-29T14:48:01.933000", "content": "Thanks, i will check again
", "votes": 0, "replies": [] } ] } ] }, { "id": 2539766, "author_name": "hengck23", "author_url": "", "post_date": "2023-11-27T09:53:46.590000", "content": "
use 3d connected components to filter false positive errors
\nhttps://pypi.org/project/connected-components-3d/
In your notebook https://www.kaggle.com/code/hengck23/lb0-808-resnet50-2d-unet-xy-zy-zx-cc3d you didn't use connected-components-3d:
\ncc_threshold = -1
\nWhat were your results using cc3d?
useful:
\nhttps://github.com/JacobBumgarner/VesselVio/tree/main
\nVesselVio is an open-source application designed for the analysis and visualization of segmented vasculature datasets.
https://blog.research.google/2021/06/a-browsable-petascale-reconstruction-of.html
\nthis competition reminds me of the ancient google paper on flood filling network (FFN)
grow_mask = \n
\n
\n
\nHigh-precision automated reconstruction of neurons with flood-filling networks
\nhttps://www.nature.com/articles/s41592-018-0049-4
\nhttps://www.biorxiv.org/content/10.1101/200675v1.full
\nhttps://www.mpg.de/12130750/neural-networks-connectome
\nhttps://www.youtube.com/watch?v=46SksPonI8I
GCO post processing:
\n(Global Constructive Optimization algorithm for generating smaller vessels)
\n[1] A Hybrid Approach to Full-Scale Reconstruction of Renal Arterial Network
\nhttps://arxiv.org/pdf/2303.01837.pdf
(e) is results from your deep network.
\n(j) is the results of tree growing with GCO
proposed solution:
\n
are you planning to apply a patch based approach? or taking the whole image into account? will we be able to fit the whole image into memory with the kaggle infrastructure?
", "votes": 1, "replies": [] } ] }, { "id": 2560267, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-13T12:29:28.990000", "content": "useful
\nhttps://github.com/PengyiZhang/RegionGrowth
\nhttps://github.com/ylmzkaan/3DRegionGrowing
\nhttp://notmatthancock.github.io/2017/10/09/region-growing-wrapping-c.html
high resolution 2d/3d unet solution !!!
\ndemo notebook :https://www.kaggle.com/code/hengck23/2d-to-3d-unet-demo
\n
\n
how to properly design 3d solution.
\nif you cannot fit a single 3d net work (because of memory GPU constraint),
\nyou properly need at least 2 net (of different hierarchy scale).
posting it here for visibility Patch based 3D UNET + TF + TPU: https://www.kaggle.com/code/dhinkris/training-patch-based-3d-unet-tf-tpu
", "votes": 1, "replies": [ { "id": 2557807, "author_name": "hengck23", "author_url": "", "post_date": "2023-12-11T17:15:07.697000", "content": "you should try these pretrain 3d unet
\nhttps://github.com/Tencent/MedicalNet