{ "id": 519628, "title": "[placeholder] my experimental results", "url": "/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/519628", "author_name": "hengck23", "post_date": "2024-07-12T05:46:17.263000", "votes": 124, "comment_count": 42, "views": 0, "content": "

… to be updated …
\nbut first, let's study some reference paper:
\n[1] DEEP SPINE: AUTOMATED LUMBAR VERTEBRAL SEGMENTATION, DISC-LEVEL DESIGNATION, AND SPINAL STENOSIS GRADING USING DEEP LEARNING
\nhttps://proceedings.mlr.press/v85/lu18a/lu18a.pdf

\n

hint: expand by searching for paper that references or referenced by this paper

\n

[2] Deep Learning Model for Automated Detection and Classification of Central Canal, Lateral Recess, and NeuralForaminal Stenosis at Lumbar Spine MRI
\nhttps://pubs.rsna.org/doi/epdf/10.1148/radiol.2021204289
\ncode: https://github.com/NUHS-NUS-SpineAI/SpineAI-Detect-Classify-LumbarMRI-Stenosis

", "messages": [ { "id": 2918256, "postDate": "2024-07-12T05:46:17.263Z", "content": "

… to be updated …
\nbut first, let's study some reference paper:
\n[1] DEEP SPINE: AUTOMATED LUMBAR VERTEBRAL SEGMENTATION, DISC-LEVEL DESIGNATION, AND SPINAL STENOSIS GRADING USING DEEP LEARNING
\nhttps://proceedings.mlr.press/v85/lu18a/lu18a.pdf

\n

hint: expand by searching for paper that references or referenced by this paper

\n

[2] Deep Learning Model for Automated Detection and Classification of Central Canal, Lateral Recess, and NeuralForaminal Stenosis at Lumbar Spine MRI
\nhttps://pubs.rsna.org/doi/epdf/10.1148/radiol.2021204289
\ncode: https://github.com/NUHS-NUS-SpineAI/SpineAI-Detect-Classify-LumbarMRI-Stenosis

", "rawMarkdown": " ... to be updated ...\nbut first, let's study some reference paper:\n\n[1] DEEP SPINE: AUTOMATED LUMBAR VERTEBRAL SEGMENTATION, DISC-LEVEL DESIGNATION, AND SPINAL STENOSIS GRADING USING DEEP LEARNING\nhttps://proceedings.mlr.press/v85/lu18a/lu18a.pdf\n\nhint: expand by searching for paper that references or referenced by this paper\n\n---\n\n[2] Deep Learning Model for Automated Detection and Classification of Central Canal, Lateral Recess, and NeuralForaminal Stenosis at Lumbar Spine MRI\nhttps://pubs.rsna.org/doi/epdf/10.1148/radiol.2021204289\ncode: https://github.com/NUHS-NUS-SpineAI/SpineAI-Detect-Classify-LumbarMRI-Stenosis", "votes": 123 }, { "id": 2918264, "postDate": "2024-07-12T05:48:15.950Z", "content": "

you need that this paper[1] is very close to what we are doing
\ndemo video: https://drive.google.com/file/d/1wnV757LRAnw9ZbVd8Xy3zUMC_i-eJtfw/view

\n

\n

\n

\n

", "rawMarkdown": "you need that this paper[1] is very close to what we are doing\ndemo video: https://drive.google.com/file/d/1wnV757LRAnw9ZbVd8Xy3zUMC_i-eJtfw/view\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F8cc91a84f6f28762f2b07a4134ff2c34%2FSelection_055.png?generation=1720763243996636&alt=media)\n\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fda6fcb82bad11e0297b8764f5d153e53%2FSelection_052.png?generation=1720763258002233&alt=media)\n\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F2d2ca597d92dc3c379c59d031907fe8b%2FSelection_053.png?generation=1720763272189615&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F325f927c3eea3601a5b8cbdb58a81798%2FSelection_054.png?generation=1720763284714296&alt=media)", "votes": 14, "replies": [ { "id": 2943730, "postDate": "2024-08-01T19:55:40.843Z", "content": "

\"you need that this paper[1] is very close to what we are doing\" Could you give the full paper name?

", "rawMarkdown": "\"you need that this paper[1] is very close to what we are doing\" Could you give the full paper name?", "replies": [ { "id": 2943976, "postDate": "2024-08-02T03:48:51.727Z", "content": "

You can search for it on Google as DeepSpine research paper

", "rawMarkdown": "You can search for it on Google as DeepSpine research paper" }, { "id": 2944420, "postDate": "2024-08-02T12:22:27.083Z", "content": "

DEEP SPINE: AUTOMATED LUMBAR VERTEBRAL SEGMENTATION, DISC-LEVEL DESIGNATION, AND SPINAL STENOSIS GRADING USING DEEP LEARNING https://www.semanticscholar.org/reader/7943d26b3ef92849dc625428e56e06f37608648c

", "rawMarkdown": "DEEP SPINE: AUTOMATED LUMBAR VERTEBRAL SEGMENTATION, DISC-LEVEL DESIGNATION, AND SPINAL STENOSIS GRADING USING DEEP LEARNING [https://www.semanticscholar.org/reader/7943d26b3ef92849dc625428e56e06f37608648c](url)", "votes": 1 } ] } ] }, { "id": 2919686, "postDate": "2024-07-13T06:04:42.047Z", "content": "

HERE is the plan!

\n

After reading many papers and consider the GPU resource for code competition(p100 or 2xT4), this is
\n[plan-A]: multivew two-stage rpn-frcnn net

\n\n

\n

[plan-B] exit plan

\n\n

[plan-C] exit plan

\n", "rawMarkdown": "HERE is the plan!\n\nAfter reading many papers and consider the GPU resource for code competition(p100 or 2xT4), this is\n[plan-A]: multivew two-stage rpn-frcnn net\n\n- joint treat objects as points. directly regress on input images to give (x,y,level,condition) objects as output for each view\n- we can use kaggle label csv directly for training.\n- these are proposals\n- then crop 3d volume for verification for each proposal\n- finally aggregate to study level\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fe0ee27a00d4956943ab75d52193c12df%2FSelection_073.png?generation=1720850421846598&alt=media)\n\n[plan-B] exit plan\n- detect all level (x,y) (we need external data for this because we need x,y for all level in train)\n- then crop 3d volume for each level to predict condition\n- finally aggregate to study level\n\n[plan-C] exit plan\n- make heatmap level(we can use kaggle data only or/and external data in train)\n- we do not predict x,y. but instead, the heatmap are additional input channel to a final image/volume classifier. \n- final image/volume (or image seq) classifier. ", "votes": 9 }, { "id": 2919438, "postDate": "2024-07-12T21:24:31.567Z", "content": "

how to read MRI:

\n

search keyword: Lumbar Spinal Stenosis, Lumbar MRI, beginner
\nhttps://www.youtube.com/watch?v=feHB0mGnpBs
\nhttps://www.youtube.com/watch?v=FkxvTsfDQ0Y
\nhttps://www.youtube.com/watch?v=qYSG2KXLNyU
\nhttps://www.youtube.com/watch?v=A9InC8ppa9k
\nhttps://www.youtube.com/watch?v=ZtehC7q_DWc
\nhttps://www.youtube.com/watch?v=qivyCxaf2IY

\n

\n

\n

\n

medical knowledge:
\nSpine Anatomy | Know Your Spine
\nhttps://www.youtube.com/watch?v=gUG_zbKqlaU
\nhttps://www.youtube.com/watch?v=O-zR_YRJNhM
\nhttps://www.youtube.com/watch?v=YtAk1GxtlUk
\nhttps://www.youtube.com/watch?v=aQyhj9V8UxE

", "rawMarkdown": "how to read MRI:\n\nsearch keyword: Lumbar Spinal Stenosis, Lumbar MRI, beginner\nhttps://www.youtube.com/watch?v=feHB0mGnpBs\nhttps://www.youtube.com/watch?v=FkxvTsfDQ0Y\nhttps://www.youtube.com/watch?v=qYSG2KXLNyU\nhttps://www.youtube.com/watch?v=A9InC8ppa9k\nhttps://www.youtube.com/watch?v=ZtehC7q_DWc\nhttps://www.youtube.com/watch?v=qivyCxaf2IY\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F2cbd3dfc8a86041d17958612129e4b56%2FSelection_069.png?generation=1720822165743461&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fc36bfbc36e79197078b6427e2c0ada19%2FSelection_068.png?generation=1720819461134224&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F5b16d86437040adecd47457e436e670c%2FSelection_070.png?generation=1720822363328885&alt=media)\n\nmedical knowledge:\nSpine Anatomy | Know Your Spine\nhttps://www.youtube.com/watch?v=gUG_zbKqlaU\nhttps://www.youtube.com/watch?v=O-zR_YRJNhM\nhttps://www.youtube.com/watch?v=YtAk1GxtlUk\nhttps://www.youtube.com/watch?v=aQyhj9V8UxE", "votes": 8 }, { "id": 2922234, "postDate": "2024-07-15T03:07:30.733Z", "content": "

good news. initial resnet18 proposal network is successful.
\ncode to release later …

\n

left to right: input, truth, predict

\n

\n

", "rawMarkdown": "good news. initial resnet18 proposal network is successful.\ncode to release later ...\n\nleft to right: input, truth, predict\n\n ![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F03c12bffb744fc54a20fa8c0e18f349d%2FSelection_094.png?generation=1721035171133361&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F178b364f42f66b3b46c018ccd171093b%2FSelection_092.png?generation=1721035187259184&alt=media)", "votes": 5 }, { "id": 2927495, "postDate": "2024-07-18T14:03:43.393Z", "content": "

This discussion post is a gemstone. Definitely going to spend quality time on it.

", "rawMarkdown": "This discussion post is a gemstone. Definitely going to spend quality time on it. ", "votes": 3 }, { "id": 2923507, "postDate": "2024-07-16T00:30:43.757Z", "content": "

yet another out of box solution:
\nhttps://github.com/rwindsor1/SpineNet
\n

", "rawMarkdown": "yet another out of box solution:\nhttps://github.com/rwindsor1/SpineNet\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F4869f5d1fbb881b41e7d3024e1e1dea6%2FSelection_105.png?generation=1721089841498777&alt=media)", "votes": 4, "replies": [ { "id": 2930953, "postDate": "2024-07-21T14:43:35.440Z", "content": "

Do you also have SpineNetv2, or do you think that would be useful?

", "rawMarkdown": "Do you also have SpineNetv2, or do you think that would be useful?", "replies": [ { "id": 2933263, "postDate": "2024-07-23T15:33:09.717Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 2933265, "postDate": "2024-07-23T15:33:36.403Z", "content": "

the licensing is weird. it is non-commercial with some caveats.

", "rawMarkdown": "the licensing is weird. it is non-commercial with some caveats.", "votes": 3 } ] }, { "id": 2918318, "postDate": "2024-07-12T06:29:38.973Z", "content": "

how to use point annotation:
\n[5] A New Window Loss Function for Bone Fracture Detection and Localization in X-ray Images with Point-based Annotation
\nhttps://www.semanticscholar.org/reader/a5ac9aca8ea61b0126729a602e91b3886e345887

\n

\n

", "rawMarkdown": "how to use point annotation:\n[5] A New Window Loss Function for Bone Fracture Detection and Localization in X-ray Images with Point-based Annotation\nhttps://www.semanticscholar.org/reader/a5ac9aca8ea61b0126729a602e91b3886e345887\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fea23e16a6a9a3a5d2e438b3c7865fae0%2FSelection_059.png?generation=1720765762146035&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fae787ea6e8b33eca02b0e1b703af1b5f%2FSelection_060.png?generation=1720765776587799&alt=media)", "votes": 4, "replies": [ { "id": 2919547, "postDate": "2024-07-13T02:04:55.960Z", "content": "

this is maybe a better option:
\nhttps://arxiv.org/abs/1904.07850

\n

paper: Objects as Points
\n\"Our detector uses keypoint estimation to find center points and regresses to all other object properties, such as size, 3D location, orientation, and even pose.\"

", "rawMarkdown": "this is maybe a better option:\nhttps://arxiv.org/abs/1904.07850\n\npaper: Objects as Points\n\"Our detector uses keypoint estimation to find center points and regresses to all other object properties, such as size, 3D location, orientation, and even pose.\"" } ] }, { "id": 2919954, "postDate": "2024-07-13T10:50:27.050Z", "content": "

commerical product: RadiSpine
\nhttps://www.youtube.com/watch?v=F5e9uYc0xHs
\nhttps://www.radirad.com/radispine-features

", "rawMarkdown": "commerical product: RadiSpine\nhttps://www.youtube.com/watch?v=F5e9uYc0xHs\nhttps://www.radirad.com/radispine-features", "votes": 1, "replies": [ { "id": 2919962, "postDate": "2024-07-13T11:01:18.800Z", "content": "

out of box solution:

\n\n

dataset:https://spider.grand-challenge.org/data/
\nhttps://www.nature.com/articles/s41597-024-03090-w

\n

\"The original training code of the IIS baseline algorithm and the trained weights and biases are publicly available at: https://github.com/DIAGNijmegen/SPIDER-Baseline-IIS. T\"

\n

", "rawMarkdown": "out of box solution:\n- Developing custom 3D medical image segmentation solutions using out-of-the-box pipelines in MONA\nhttps://www.kitware.com/developing-custom-3d-medical-image-segmentation-solutions-using-out-of-the-box-pipelines-in-monai/\n\ndataset:https://spider.grand-challenge.org/data/\nhttps://www.nature.com/articles/s41597-024-03090-w\n\n\"The original training code of the IIS baseline algorithm and the trained weights and biases are publicly available at: https://github.com/DIAGNijmegen/SPIDER-Baseline-IIS. T\"\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F8fcdf147a38d833d67aead333d530b66%2FSelection_074.png?generation=1720869963741102&alt=media)\n", "votes": 1, "replies": [ { "id": 2919969, "postDate": "2024-07-13T11:04:35.983Z", "content": "

inpainting as data augmentatin
\nInpainting Pathology in Lumbar Spine MRI with Latent Diffusion
\nhttps://arxiv.org/pdf/2406.02477

", "rawMarkdown": "inpainting as data augmentatin\nInpainting Pathology in Lumbar Spine MRI with Latent Diffusion\nhttps://arxiv.org/pdf/2406.02477", "votes": 1 }, { "id": 2920198, "postDate": "2024-07-13T13:53:35.877Z", "content": "

Can we use this RSNA 2024 Lumbar Spine Degenerative Classification -> this competition dataset for publishing research paper later?

", "rawMarkdown": "Can we use this RSNA 2024 Lumbar Spine Degenerative Classification -> this competition dataset for publishing research paper later?\n" } ] } ] }, { "id": 2919377, "postDate": "2024-07-12T20:07:03.693Z", "content": "

other papers that i have no time to read in details but are very much related. Maybe someone could get chatgpt to read them.

\n\n

\n\n

\n", "rawMarkdown": "other papers that i have no time to read in details but are very much related. Maybe someone could get chatgpt to read them.\n\n- Deep learning-based high-accuracy quantitation for lumbar intervertebral disc degeneration from MRI\nhttps://www.nature.com/articles/s41467-022-28387-5\ncode: https://github.com/no-saint-no-angel/BianqueNet\n\n- Deep learning-based high-accuracy detection for lumbar and cervical degenerative disease on T2-weighted MR images\nhttps://www.researchgate.net/publication/369413726_Deep_learning-based_high-accuracy_detection_for_lumbar_and_cervical_degenerative_disease_on_T2-weighted_MR_images\n(3D ResNet18 and transformer(multi-view fusion))\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F983c6a251387a6b392b0a95ab318f1a2%2FSelection_064.png?generation=1720814884240351&alt=media)\n\n- Context-Aware Transformers For Spinal Cancer Detection and Radiological Grading\nhttps://arxiv.org/pdf/2206.13173\n(2D ResNet18 + seq + patch and transformer(multi-view fusion))\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ff2e701a22f1f5e9ca4f46c94f019cdb4%2FSelection_067.png?generation=1720815606424594&alt=media)\n\n\n- A deep learning model for detection of cervical spinal cord compression in MRI scans\nhttps://www.nature.com/articles/s41598-021-89848-3\n(resnet50 CNN, only axial images)\n", "votes": 2, "replies": [ { "id": 2919380, "postDate": "2024-07-12T20:14:04.873Z", "content": "", "rawMarkdown": "- Improving Trustworthiness of AI Disease Severity Rating in Medical Imaging with Ordinal Conformal Prediction Sets\nhttps://arxiv.org/pdf/2207.02238\ncode here: https://github.com/clu5/lumbar-conformal\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa011430e2733d1105aa1b25c30f6b6cb%2FSelection_066.png?generation=1720815242492221&alt=media)", "votes": 1 } ] }, { "id": 2918304, "postDate": "2024-07-12T06:21:10.873Z", "content": "

[1] SpineOne: A One-Stage Detection Framework for Degenerative Discs and Vertebrae
\nhttps://arxiv.org/pdf/2110.15082

\n

this work uses keypoint:
\n\"we propose a one-stage detection framework termed SpineOne to simultaneously localize and classify degenerative discs and vertebrae from MRI slices. SpineOne is built upon the following three key techniques: 1) a new design of the keypoint heatmap to facilitate simultaneous keypoint localization and classification; 2) …\"

\n

", "rawMarkdown": "[1] SpineOne: A One-Stage Detection Framework for Degenerative Discs and Vertebrae\nhttps://arxiv.org/pdf/2110.15082\n\nthis work uses keypoint:\n\"we propose a one-stage detection framework termed SpineOne to simultaneously localize and classify degenerative discs and vertebrae from MRI slices. SpineOne is built upon the following three key techniques: 1) a new design of the keypoint heatmap to facilitate simultaneous keypoint localization and classification; 2) ...\"\n\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fd2c855964ef9e449b43ea530514ff706%2FSelection_058.png?generation=1720765268300160&alt=media)", "votes": 2, "replies": [ { "id": 2930860, "postDate": "2024-07-21T13:24:54.760Z", "content": "

I have joined point detection with main model but not improve 😭 I am debugging

", "rawMarkdown": "I have joined point detection with main model but not improve 😭 I am debugging ", "votes": 1, "replies": [ { "id": 2948810, "postDate": "2024-08-06T06:26:50.247Z", "content": "

Did it work?

", "rawMarkdown": "Did it work?" }, { "id": 2949960, "postDate": "2024-08-07T05:23:23.383Z", "content": "

@quan0095 and @arrosw

\n

the trick is to go beyond the label coords given in the ground truth of the train dataset.

\n

for example, the given annotation is:

\n
\n    \n    \n    \n\n\n    \n    \n    \n
\n

now if you have many training data, we can train as normal.

\n

but there is limited data, it is difficult to differentiate between slice3 and slice4:

\n\n

so it is an easier problem to train (model input=one slice) if we use:

\n
\n    \n    \n    \n\n\n    \n    \n    \n
\n
\n

alternative,model input =[multiple slice]:

\n
def model_forward(slice3,slice4):\n     e3 = encode(slice3)\n     e4 = encode(slice4)\n     pool =  pooling(e3,e4)\n    x,y,grade = classifier(pool)\n     # one common   grade predicted  N  slices\n
", "rawMarkdown": "@quan0095 and @arrosw \n\nthe trick is to go beyond the label coords given in the ground truth of the train dataset.\n\nfor example, the given annotation is:\n\n```\nslice3:\n condition_l1_l2 : x,y,grade\n condition_l2_l3 : x,y,grade\n condition_l3_l4 : missing\n\nslice4 (nearby slice):\n condition_l1_l2 : missing\n condition_l2_l3 : missing\n condition_l3_l4 : x,y,grade\n\n```\n\nnow if you have many training data, we can train as normal.\n\nbut there is limited data, it is difficult to differentiate between slice3 and slice4:\n- slice 3 and slice4 are very similar\n- our target is NOT to predict the grading for each slide\n- rather we want to predict the one grading for [slide3, slide4]\n\nso it is an easier problem to train (model input=one slice) if we use:\n```\nslice3:\n condition_l1_l2 : x,y,grade\n condition_l2_l3 : x,y,grade\n condition_l3_l4 : copy from slice4\n\nslice4 (nearby slice):\n condition_l1_l2 : copy from slice3\n condition_l2_l3 : copy from slice3\n condition_l3_l4 : x,y,grade\n\n```\n\n---\n\nalternative,model input =[multiple slice]:\n\n```\ndef model_forward(slice3,slice4):\n e3 = encode(slice3)\n e4 = encode(slice4)\n pool = some pooling(e3,e4)\n x,y,grade = classifier(pool)\n #only one common point and grade predicted for N input slices\n\n```", "votes": 2 }, { "id": 2954387, "postDate": "2024-08-09T16:44:21.127Z", "content": "

Hello,
\nHow do you select interval around each slice from which it can receive annotations from other slices if not present for itself?

", "rawMarkdown": "Hello,\nHow do you select interval around each slice from which it can receive annotations from other slices if not present for itself?" } ] } ] }, { "id": 2918272, "postDate": "2024-07-12T05:57:09.333Z", "content": "

paper[2] is an object detection based method using faster-rcnn

\n

\n

", "rawMarkdown": "paper[2] is an object detection based method using faster-rcnn\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa38ee434344d1b4114172ce13e10e265%2FSelection_056.png?generation=1720763809013090&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fe2f2a409f0ac14b36483e309d7b356b9%2FSelection_057.png?generation=1720763824492488&alt=media)\n", "votes": 2, "replies": [ { "id": 2919309, "postDate": "2024-07-12T19:25:06.763Z", "content": "

related: using mask-rcnn
\npaper: Deep Learning-Based Intelligent Diagnosis of Lumbar Diseases with Multi-Angle View of Intervertebral Disc
\nhttps://www.mdpi.com/2227-7390/12/13/2062

\n

external dataset:
\nhttps://tianchi.aliyun.com/dataset/79463 (CC-BY-SA-NC 4.0)
\nhttps://tianchi.aliyun.com/competition/entrance/531796/information
\n(check the forum)

\n

The dataset includes MRI images of T1 and T2 sagittal plane and T2 axial plane (FSE/TSE).

\n

\n

dataset annotation

\n

\n
\n

WRANING!!!! Please check the dataset usgae copyright yourself!!!!

\n
\n

hnint : google for \"脊柱 阿里云天池\" , \"Spark“数字人体”AI挑战赛\", for code, paper, tutorial, etc
\ne.g. code: https://github.com/wolaituodiban/spinal_detection_baseline

\n

脊柱疾病智能诊断-GPU赛道-deep thinker-冠军比赛方案
\nIntelligent diagnosis of spinal diseases-GPU track: deep thinker winner solution
\nhttps://tianchi.aliyun.com/course/live/1541

\n

https://tianchi.aliyun.com/forum/post/135949?spm=a2c22.28136470.0.0.20b37a0bFhMfYD&from=search-list
\n(same as the paper?)

\n

spine alignment trick
\n

", "rawMarkdown": "related: using mask-rcnn\npaper: Deep Learning-Based Intelligent Diagnosis of Lumbar Diseases with Multi-Angle View of Intervertebral Disc\nhttps://www.mdpi.com/2227-7390/12/13/2062\n\nexternal dataset:\nhttps://tianchi.aliyun.com/dataset/79463 (CC-BY-SA-NC 4.0)\nhttps://tianchi.aliyun.com/competition/entrance/531796/information\n(check the forum)\n\nThe dataset includes MRI images of T1 and T2 sagittal plane and T2 axial plane (FSE/TSE).\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F5cf21295fb55998b41c0eedc1dad6880%2FSelection_061.png?generation=1720812577435335&alt=media)\n\n\ndataset annotation\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F6f9094b4d034ab3c9d6ca82c9fa1a0d3%2FSelection_063.png?generation=1720814093147105&alt=media)\n\n---\n\n**WRANING!!!! Please check the dataset usgae copyright yourself!!!!**\n\n---\n\nhnint : google for \"脊柱 阿里云天池\" , \"Spark“数字人体”AI挑战赛\", for code, paper, tutorial, etc\ne.g. code: https://github.com/wolaituodiban/spinal_detection_baseline\n\n脊柱疾病智能诊断-GPU赛道-deep thinker-冠军比赛方案\nIntelligent diagnosis of spinal diseases-GPU track: deep thinker winner solution\nhttps://tianchi.aliyun.com/course/live/1541\n\nhttps://tianchi.aliyun.com/forum/post/135949?spm=a2c22.28136470.0.0.20b37a0bFhMfYD&from=search-list\n(same as the paper?)\n\nspine alignment trick \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F3d1fbad8ef5c79f3ff1797ead98c0a56%2FSelection_062.png?generation=1720813474678620&alt=media)", "votes": 4, "replies": [ { "id": 2919395, "postDate": "2024-07-12T20:29:48.693Z", "content": "

another related external dataset:
\nhttps://ivdm3seg.weebly.com/
\nMICCAI 2018 Challenge Automatic Intervertebral Disc Localization and Segmentation from 3D Multi-modality MR (M3) Images

\n

more here: http://spineweb.digitalimaginggroup.ca/Index.php?n=Main.Datasets

\n
\n

not sure if this is useful:

\n", "rawMarkdown": "another related external dataset:\nhttps://ivdm3seg.weebly.com/\nMICCAI 2018 Challenge Automatic Intervertebral Disc Localization and Segmentation from 3D Multi-modality MR (M3) Images \n\nmore here: http://spineweb.digitalimaginggroup.ca/Index.php?n=Main.Datasets\n\n---\nnot sure if this is useful:\n- Reproducible Spinal Cord Quantitative MRI Analysis with the Spinal Cord Toolbox\nhttps://www.jstage.jst.go.jp/article/mrms/23/3/23_rev.2023-0159/_pdf/-char/en\nhttps://github.com/spinalcordtoolbox/spinalcordtoolbox", "votes": 1 }, { "id": 2919674, "postDate": "2024-07-13T05:36:03.527Z", "content": "


\nhttps://tianchi.aliyun.com/course/live/1541

\n

Introduction
\nSpark “Digital Human” AI Challenge

\n

——A total of 12 teams participated in the final of the Spinal Disease Intelligent Diagnosis Competition, divided into two tracks: GPU and CPU. The final results are as follows:

\n

GPU track champion deep thinker
\nGPU track runner-up triple-Z, shiontao
\nThe third runner-up in the GPU track is Zhejiang University Ruiyi, Xiaotuanzi, I am a bricklayer

\n

CPU track champion sailor
\nCPU track runner-up U-net, 2:30 p.m.
\nThird runner-up in the CPU track AI explorer, Tangguo, just run around

", "rawMarkdown": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fd3bd11c87074f94eed4a666f9259acb2%2FSelection_071.png?generation=1720848948691077&alt=media)\nhttps://tianchi.aliyun.com/course/live/1541\n\nIntroduction\nSpark “Digital Human” AI Challenge\n\n——A total of 12 teams participated in the final of the Spinal Disease Intelligent Diagnosis Competition, divided into two tracks: GPU and CPU. The final results are as follows:\n\nGPU track champion deep thinker\nGPU track runner-up triple-Z, shiontao\nThe third runner-up in the GPU track is Zhejiang University Ruiyi, Xiaotuanzi, I am a bricklayer\n\nCPU track champion sailor\nCPU track runner-up U-net, 2:30 p.m.\nThird runner-up in the CPU track AI explorer, Tangguo, just run around\n", "votes": 1 }, { "id": 2981893, "postDate": "2024-09-07T08:24:31.617Z", "content": "

thanks for sharing so much useful information!

\n

Is it OK to use the coord label in this dataset?
\nhttps://tianchi.aliyun.com/dataset/79463 (CC-BY-SA-NC 4.0)
\nThe license is OK, but I find I need an alibaba cloud account to download the data, thus I am confused whether it is ok to use the dataset in Kaggle. What do you think?

", "rawMarkdown": "thanks for sharing so much useful information!\n\nIs it OK to use the coord label in this dataset?\nhttps://tianchi.aliyun.com/dataset/79463 (CC-BY-SA-NC 4.0)\nThe license is OK, but I find I need an alibaba cloud account to download the data, thus I am confused whether it is ok to use the dataset in Kaggle. What do you think?\n", "votes": 1 } ] } ] }, { "id": 3003238, "postDate": "2024-09-30T19:17:58.793Z", "content": "

Both research papers are having great insights. Thank you for sharing.
\nI have few questions about it, please help me to understand following.

\n
    \n

  1. As model is works on 2 stages, ROI detection and classification. How and what to provide I/p( like which instance of mri image) to Roi detection. As for orientation sagital there multiple slices and only few slices has clear visibility of disc?

    2. \n

  2. As in paper one after detection of verbetrals cord and disc, each disc images get detected as forms 3d volume and their respective axial orientation images for sub-articular stenosis. how each disc level relation is getting establish for axial and sagital?

    3. \n
", "rawMarkdown": "Both research papers are having great insights. Thank you for sharing.\nI have few questions about it, please help me to understand following.\n1. As model is works on 2 stages, ROI detection and classification. How and what to provide I/p( like which instance of mri image) to Roi detection. As for orientation sagital there multiple slices and only few slices has clear visibility of disc?\n\n2. As in paper one after detection of verbetrals cord and disc, each disc images get detected as forms 3d volume and their respective axial orientation images for sub-articular stenosis. how each disc level relation is getting establish for axial and sagital?" }, { "id": 2928181, "postDate": "2024-07-19T02:14:28.767Z", "content": "

Regarding the multi-input, mult-task, multi-class model described in paper [1], how does one structure the target labels such that we have both condition at disc level such as left_foraminal_l1/l2 (multi-task) and the severity classes (multi-class)?

", "rawMarkdown": "Regarding the multi-input, mult-task, multi-class model described in paper [1], how does one structure the target labels such that we have both condition at disc level such as left_foraminal_l1/l2 (multi-task) and the severity classes (multi-class)? ", "replies": [ { "id": 2928192, "postDate": "2024-07-19T02:29:18.353Z", "content": "

If you are talking about the DeepSpine paper, their pipeline is pretty much: use segmentation from saggital view to split the volume to five levels, that's multi-class. For each level, they have both axial and sagittal input, that's multi-input. and finally, like us, they predict four or three severity labels for each degeneration on each level. The hard part would be the segmentation as we need extra data to train it.

", "rawMarkdown": "If you are talking about the DeepSpine paper, their pipeline is pretty much: use segmentation from saggital view to split the volume to five levels, that's multi-class. For each level, they have both axial and sagittal input, that's multi-input. and finally, like us, they predict four or three severity labels for each degeneration on each level. The hard part would be the segmentation as we need extra data to train it.", "votes": 4, "replies": [ { "id": 2928194, "postDate": "2024-07-19T02:31:47.783Z", "content": "

have you considered SPIDER dataset for segmentation part ? -https://zenodo.org/records/10159290

", "rawMarkdown": "have you considered SPIDER dataset for segmentation part ? -[https://zenodo.org/records/10159290](url)", "votes": 1 }, { "id": 2928199, "postDate": "2024-07-19T02:37:12.303Z", "content": "

I am following this post's step to train a object-as-point multi-channel prediction for each level, because it directly use the labels coordinates from the competition. If that doesn't work, I think SPIDER might be very helpful

", "rawMarkdown": "I am following this post's step to train a object-as-point multi-channel prediction for each level, because it directly use the labels coordinates from the competition. If that doesn't work, I think SPIDER might be very helpful" } ] } ] }, { "id": 2923531, "postDate": "2024-07-16T00:52:20.170Z", "content": "

Great! Keep going! Looking forward for more coming today

", "rawMarkdown": "Great! Keep going! Looking forward for more coming today" }, { "id": 2922525, "postDate": "2024-07-15T09:01:35.553Z", "content": "

Thank you hengck so much for providing so many useful papers. I always admire ability to find amazing papers. May I ask where did you find them? do you have a specific place to go to and/or method for finding those related papers?

", "rawMarkdown": "Thank you hengck so much for providing so many useful papers. I always admire ability to find amazing papers. May I ask where did you find them? do you have a specific place to go to and/or method for finding those related papers? ", "replies": [ { "id": 2922541, "postDate": "2024-07-15T09:22:20.467Z", "content": "

@llleeeoooh
\nit is about using the correct keyword. I just use google.

", "rawMarkdown": "@llleeeoooh \nit is about using the correct keyword. I just use google.", "votes": 4 } ] }, { "id": 2922469, "postDate": "2024-07-15T08:03:15.463Z", "content": "

I have been trying to replicate Deep SPINE research work. It has good intuition for addressing multi-input multi-task and multi-class classification.
\nIf, anyone else is also trying to work on this, kindly get in touch. Looking for a teammate!

", "rawMarkdown": "I have been trying to replicate Deep SPINE research work. It has good intuition for addressing multi-input multi-task and multi-class classification.\nIf, anyone else is also trying to work on this, kindly get in touch. Looking for a teammate! \n" }, { "id": 2918811, "postDate": "2024-07-12T14:15:53.007Z", "content": "

awesome papers!! gonna check them out

", "rawMarkdown": "awesome papers!! gonna check them out" }, { "id": 3013232, "postDate": "2024-10-09T19:06:18.847Z", "content": "

Very helpful! Thanks

", "rawMarkdown": "Very helpful! Thanks" }, { "id": 2920545, "postDate": "2024-07-13T17:21:52.697Z", "content": "

lot of valuable info. Thanks.

", "rawMarkdown": "lot of valuable info. Thanks." } ], "comments": [ { "id": 2918264, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-12T05:48:15.950000", "content": "

you need that this paper[1] is very close to what we are doing
\ndemo video: https://drive.google.com/file/d/1wnV757LRAnw9ZbVd8Xy3zUMC_i-eJtfw/view

\n

\n

\n

\n

", "votes": 14, "replies": [ { "id": 2943730, "author_name": "DrDC", "author_url": "", "post_date": "2024-08-01T19:55:40.843000", "content": "

\"you need that this paper[1] is very close to what we are doing\" Could you give the full paper name?

", "votes": 0, "replies": [ { "id": 2943976, "author_name": "Rahul Nakka", "author_url": "", "post_date": "2024-08-02T03:48:51.727000", "content": "

You can search for it on Google as DeepSpine research paper

", "votes": 0, "replies": [] }, { "id": 2944420, "author_name": "DrDC", "author_url": "", "post_date": "2024-08-02T12:22:27.083000", "content": "

DEEP SPINE: AUTOMATED LUMBAR VERTEBRAL SEGMENTATION, DISC-LEVEL DESIGNATION, AND SPINAL STENOSIS GRADING USING DEEP LEARNING https://www.semanticscholar.org/reader/7943d26b3ef92849dc625428e56e06f37608648c

", "votes": 1, "replies": [] } ] } ] }, { "id": 2919686, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-13T06:04:42.047000", "content": "

HERE is the plan!

\n

After reading many papers and consider the GPU resource for code competition(p100 or 2xT4), this is
\n[plan-A]: multivew two-stage rpn-frcnn net

\n\n

\n

[plan-B] exit plan

\n\n

[plan-C] exit plan

\n", "votes": 9, "replies": [] }, { "id": 2919438, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-12T21:24:31.567000", "content": "

how to read MRI:

\n

search keyword: Lumbar Spinal Stenosis, Lumbar MRI, beginner
\nhttps://www.youtube.com/watch?v=feHB0mGnpBs
\nhttps://www.youtube.com/watch?v=FkxvTsfDQ0Y
\nhttps://www.youtube.com/watch?v=qYSG2KXLNyU
\nhttps://www.youtube.com/watch?v=A9InC8ppa9k
\nhttps://www.youtube.com/watch?v=ZtehC7q_DWc
\nhttps://www.youtube.com/watch?v=qivyCxaf2IY

\n

\n

\n

\n

medical knowledge:
\nSpine Anatomy | Know Your Spine
\nhttps://www.youtube.com/watch?v=gUG_zbKqlaU
\nhttps://www.youtube.com/watch?v=O-zR_YRJNhM
\nhttps://www.youtube.com/watch?v=YtAk1GxtlUk
\nhttps://www.youtube.com/watch?v=aQyhj9V8UxE

", "votes": 8, "replies": [] }, { "id": 2922234, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-15T03:07:30.733000", "content": "

good news. initial resnet18 proposal network is successful.
\ncode to release later …

\n

left to right: input, truth, predict

\n

\n

", "votes": 5, "replies": [] }, { "id": 2927495, "author_name": "samu2505", "author_url": "", "post_date": "2024-07-18T14:03:43.393000", "content": "

This discussion post is a gemstone. Definitely going to spend quality time on it.

", "votes": 3, "replies": [] }, { "id": 2923507, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-16T00:30:43.757000", "content": "

yet another out of box solution:
\nhttps://github.com/rwindsor1/SpineNet
\n

", "votes": 4, "replies": [ { "id": 2930953, "author_name": "Deniz Can Elci", "author_url": "", "post_date": "2024-07-21T14:43:35.440000", "content": "

Do you also have SpineNetv2, or do you think that would be useful?

", "votes": 0, "replies": [ { "id": 2933263, "author_name": "", "author_url": "", "post_date": "2024-07-23T15:33:09.717000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 2933265, "author_name": "SSS", "author_url": "", "post_date": "2024-07-23T15:33:36.403000", "content": "

the licensing is weird. it is non-commercial with some caveats.

", "votes": 3, "replies": [] } ] }, { "id": 2918318, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-12T06:29:38.973000", "content": "

how to use point annotation:
\n[5] A New Window Loss Function for Bone Fracture Detection and Localization in X-ray Images with Point-based Annotation
\nhttps://www.semanticscholar.org/reader/a5ac9aca8ea61b0126729a602e91b3886e345887

\n

\n

", "votes": 4, "replies": [ { "id": 2919547, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-13T02:04:55.960000", "content": "

this is maybe a better option:
\nhttps://arxiv.org/abs/1904.07850

\n

paper: Objects as Points
\n\"Our detector uses keypoint estimation to find center points and regresses to all other object properties, such as size, 3D location, orientation, and even pose.\"

", "votes": 0, "replies": [] } ] }, { "id": 2919954, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-13T10:50:27.050000", "content": "

commerical product: RadiSpine
\nhttps://www.youtube.com/watch?v=F5e9uYc0xHs
\nhttps://www.radirad.com/radispine-features

", "votes": 1, "replies": [ { "id": 2919962, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-13T11:01:18.800000", "content": "

out of box solution:

\n\n

dataset:https://spider.grand-challenge.org/data/
\nhttps://www.nature.com/articles/s41597-024-03090-w

\n

\"The original training code of the IIS baseline algorithm and the trained weights and biases are publicly available at: https://github.com/DIAGNijmegen/SPIDER-Baseline-IIS. T\"

\n

", "votes": 1, "replies": [ { "id": 2919969, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-13T11:04:35.983000", "content": "

inpainting as data augmentatin
\nInpainting Pathology in Lumbar Spine MRI with Latent Diffusion
\nhttps://arxiv.org/pdf/2406.02477

", "votes": 1, "replies": [] }, { "id": 2920198, "author_name": "VDevPar", "author_url": "", "post_date": "2024-07-13T13:53:35.877000", "content": "

Can we use this RSNA 2024 Lumbar Spine Degenerative Classification -> this competition dataset for publishing research paper later?

", "votes": 0, "replies": [] } ] } ] }, { "id": 2919377, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-12T20:07:03.693000", "content": "

other papers that i have no time to read in details but are very much related. Maybe someone could get chatgpt to read them.

\n\n

\n\n

\n", "votes": 2, "replies": [ { "id": 2919380, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-12T20:14:04.873000", "content": "", "votes": 1, "replies": [] } ] }, { "id": 2918304, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-12T06:21:10.873000", "content": "

[1] SpineOne: A One-Stage Detection Framework for Degenerative Discs and Vertebrae
\nhttps://arxiv.org/pdf/2110.15082

\n

this work uses keypoint:
\n\"we propose a one-stage detection framework termed SpineOne to simultaneously localize and classify degenerative discs and vertebrae from MRI slices. SpineOne is built upon the following three key techniques: 1) a new design of the keypoint heatmap to facilitate simultaneous keypoint localization and classification; 2) …\"

\n

", "votes": 2, "replies": [ { "id": 2930860, "author_name": "Quan Vu", "author_url": "", "post_date": "2024-07-21T13:24:54.760000", "content": "

I have joined point detection with main model but not improve 😭 I am debugging

", "votes": 1, "replies": [ { "id": 2948810, "author_name": "Arrosw", "author_url": "", "post_date": "2024-08-06T06:26:50.247000", "content": "

Did it work?

", "votes": 0, "replies": [] }, { "id": 2949960, "author_name": "hengck23", "author_url": "", "post_date": "2024-08-07T05:23:23.383000", "content": "

@quan0095 and @arrosw

\n

the trick is to go beyond the label coords given in the ground truth of the train dataset.

\n

for example, the given annotation is:

\n
\n    \n    \n    \n\n\n    \n    \n    \n
\n

now if you have many training data, we can train as normal.

\n

but there is limited data, it is difficult to differentiate between slice3 and slice4:

\n\n

so it is an easier problem to train (model input=one slice) if we use:

\n
\n    \n    \n    \n\n\n    \n    \n    \n
\n
\n

alternative,model input =[multiple slice]:

\n
def model_forward(slice3,slice4):\n     e3 = encode(slice3)\n     e4 = encode(slice4)\n     pool =  pooling(e3,e4)\n    x,y,grade = classifier(pool)\n     # one common   grade predicted  N  slices\n
", "votes": 2, "replies": [] }, { "id": 2954387, "author_name": "Simon Veitner", "author_url": "", "post_date": "2024-08-09T16:44:21.127000", "content": "

Hello,
\nHow do you select interval around each slice from which it can receive annotations from other slices if not present for itself?

", "votes": 0, "replies": [] } ] } ] }, { "id": 2918272, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-12T05:57:09.333000", "content": "

paper[2] is an object detection based method using faster-rcnn

\n

\n

", "votes": 2, "replies": [ { "id": 2919309, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-12T19:25:06.763000", "content": "

related: using mask-rcnn
\npaper: Deep Learning-Based Intelligent Diagnosis of Lumbar Diseases with Multi-Angle View of Intervertebral Disc
\nhttps://www.mdpi.com/2227-7390/12/13/2062

\n

external dataset:
\nhttps://tianchi.aliyun.com/dataset/79463 (CC-BY-SA-NC 4.0)
\nhttps://tianchi.aliyun.com/competition/entrance/531796/information
\n(check the forum)

\n

The dataset includes MRI images of T1 and T2 sagittal plane and T2 axial plane (FSE/TSE).

\n

\n

dataset annotation

\n

\n
\n

WRANING!!!! Please check the dataset usgae copyright yourself!!!!

\n
\n

hnint : google for \"脊柱 阿里云天池\" , \"Spark“数字人体”AI挑战赛\", for code, paper, tutorial, etc
\ne.g. code: https://github.com/wolaituodiban/spinal_detection_baseline

\n

脊柱疾病智能诊断-GPU赛道-deep thinker-冠军比赛方案
\nIntelligent diagnosis of spinal diseases-GPU track: deep thinker winner solution
\nhttps://tianchi.aliyun.com/course/live/1541

\n

https://tianchi.aliyun.com/forum/post/135949?spm=a2c22.28136470.0.0.20b37a0bFhMfYD&from=search-list
\n(same as the paper?)

\n

spine alignment trick
\n

", "votes": 4, "replies": [ { "id": 2919395, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-12T20:29:48.693000", "content": "

another related external dataset:
\nhttps://ivdm3seg.weebly.com/
\nMICCAI 2018 Challenge Automatic Intervertebral Disc Localization and Segmentation from 3D Multi-modality MR (M3) Images

\n

more here: http://spineweb.digitalimaginggroup.ca/Index.php?n=Main.Datasets

\n
\n

not sure if this is useful:

\n", "votes": 1, "replies": [] }, { "id": 2919674, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-13T05:36:03.527000", "content": "


\nhttps://tianchi.aliyun.com/course/live/1541

\n

Introduction
\nSpark “Digital Human” AI Challenge

\n

——A total of 12 teams participated in the final of the Spinal Disease Intelligent Diagnosis Competition, divided into two tracks: GPU and CPU. The final results are as follows:

\n

GPU track champion deep thinker
\nGPU track runner-up triple-Z, shiontao
\nThe third runner-up in the GPU track is Zhejiang University Ruiyi, Xiaotuanzi, I am a bricklayer

\n

CPU track champion sailor
\nCPU track runner-up U-net, 2:30 p.m.
\nThird runner-up in the CPU track AI explorer, Tangguo, just run around

", "votes": 1, "replies": [] }, { "id": 2981893, "author_name": "RihanPiggy", "author_url": "", "post_date": "2024-09-07T08:24:31.617000", "content": "

thanks for sharing so much useful information!

\n

Is it OK to use the coord label in this dataset?
\nhttps://tianchi.aliyun.com/dataset/79463 (CC-BY-SA-NC 4.0)
\nThe license is OK, but I find I need an alibaba cloud account to download the data, thus I am confused whether it is ok to use the dataset in Kaggle. What do you think?

", "votes": 1, "replies": [] } ] } ] }, { "id": 3003238, "author_name": "Rinku Sahu", "author_url": "", "post_date": "2024-09-30T19:17:58.793000", "content": "

Both research papers are having great insights. Thank you for sharing.
\nI have few questions about it, please help me to understand following.

\n
    \n

  1. As model is works on 2 stages, ROI detection and classification. How and what to provide I/p( like which instance of mri image) to Roi detection. As for orientation sagital there multiple slices and only few slices has clear visibility of disc?

    2. \n

  2. As in paper one after detection of verbetrals cord and disc, each disc images get detected as forms 3d volume and their respective axial orientation images for sub-articular stenosis. how each disc level relation is getting establish for axial and sagital?

    3. \n
", "votes": 0, "replies": [] }, { "id": 2928181, "author_name": "samu2505", "author_url": "", "post_date": "2024-07-19T02:14:28.767000", "content": "

Regarding the multi-input, mult-task, multi-class model described in paper [1], how does one structure the target labels such that we have both condition at disc level such as left_foraminal_l1/l2 (multi-task) and the severity classes (multi-class)?

", "votes": 0, "replies": [ { "id": 2928192, "author_name": "LLLEEEOOOH", "author_url": "", "post_date": "2024-07-19T02:29:18.353000", "content": "

If you are talking about the DeepSpine paper, their pipeline is pretty much: use segmentation from saggital view to split the volume to five levels, that's multi-class. For each level, they have both axial and sagittal input, that's multi-input. and finally, like us, they predict four or three severity labels for each degeneration on each level. The hard part would be the segmentation as we need extra data to train it.

", "votes": 4, "replies": [ { "id": 2928194, "author_name": "Rahul Nakka", "author_url": "", "post_date": "2024-07-19T02:31:47.783000", "content": "

have you considered SPIDER dataset for segmentation part ? -https://zenodo.org/records/10159290

", "votes": 1, "replies": [] }, { "id": 2928199, "author_name": "LLLEEEOOOH", "author_url": "", "post_date": "2024-07-19T02:37:12.303000", "content": "

I am following this post's step to train a object-as-point multi-channel prediction for each level, because it directly use the labels coordinates from the competition. If that doesn't work, I think SPIDER might be very helpful

", "votes": 0, "replies": [] } ] } ] }, { "id": 2923531, "author_name": "Syed Sajeel Haider", "author_url": "", "post_date": "2024-07-16T00:52:20.170000", "content": "

Great! Keep going! Looking forward for more coming today

", "votes": 0, "replies": [] }, { "id": 2922525, "author_name": "LLLEEEOOOH", "author_url": "", "post_date": "2024-07-15T09:01:35.553000", "content": "

Thank you hengck so much for providing so many useful papers. I always admire ability to find amazing papers. May I ask where did you find them? do you have a specific place to go to and/or method for finding those related papers?

", "votes": 0, "replies": [ { "id": 2922541, "author_name": "hengck23", "author_url": "", "post_date": "2024-07-15T09:22:20.467000", "content": "

@llleeeoooh
\nit is about using the correct keyword. I just use google.

", "votes": 4, "replies": [] } ] }, { "id": 2922469, "author_name": "Rahul Nakka", "author_url": "", "post_date": "2024-07-15T08:03:15.463000", "content": "

I have been trying to replicate Deep SPINE research work. It has good intuition for addressing multi-input multi-task and multi-class classification.
\nIf, anyone else is also trying to work on this, kindly get in touch. Looking for a teammate!

", "votes": 0, "replies": [] }, { "id": 2918811, "author_name": "JoO", "author_url": "", "post_date": "2024-07-12T14:15:53.007000", "content": "

awesome papers!! gonna check them out

", "votes": 0, "replies": [] }, { "id": 3013232, "author_name": "acitykm", "author_url": "", "post_date": "2024-10-09T19:06:18.847000", "content": "

Very helpful! Thanks

", "votes": 0, "replies": [] }, { "id": 2920545, "author_name": "dragon zhang", "author_url": "", "post_date": "2024-07-13T17:21:52.697000", "content": "

lot of valuable info. Thanks.

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2918256": " ... to be updated ...\nbut first, let's study some reference paper:\n\n[1] DEEP SPINE: AUTOMATED LUMBAR VERTEBRAL SEGMENTATION, DISC-LEVEL DESIGNATION, AND SPINAL STENOSIS GRADING USING DEEP LEARNING\nhttps://proceedings.mlr.press/v85/lu18a/lu18a.pdf\n\nhint: expand by searching for paper that references or referenced by this paper\n\n---\n\n[2] Deep Learning Model for Automated Detection and Classification of Central Canal, Lateral Recess, and NeuralForaminal Stenosis at Lumbar Spine MRI\nhttps://pubs.rsna.org/doi/epdf/10.1148/radiol.2021204289\ncode: https://github.com/NUHS-NUS-SpineAI/SpineAI-Detect-Classify-LumbarMRI-Stenosis", "2918264": "you need that this paper[1] is very close to what we are doing\ndemo video: https://drive.google.com/file/d/1wnV757LRAnw9ZbVd8Xy3zUMC_i-eJtfw/view\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F8cc91a84f6f28762f2b07a4134ff2c34%2FSelection_055.png?generation=1720763243996636&alt=media)\n\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fda6fcb82bad11e0297b8764f5d153e53%2FSelection_052.png?generation=1720763258002233&alt=media)\n\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F2d2ca597d92dc3c379c59d031907fe8b%2FSelection_053.png?generation=1720763272189615&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F325f927c3eea3601a5b8cbdb58a81798%2FSelection_054.png?generation=1720763284714296&alt=media)", "2919686": "HERE is the plan!\n\nAfter reading many papers and consider the GPU resource for code competition(p100 or 2xT4), this is\n[plan-A]: multivew two-stage rpn-frcnn net\n\n- joint treat objects as points. directly regress on input images to give (x,y,level,condition) objects as output for each view\n- we can use kaggle label csv directly for training.\n- these are proposals\n- then crop 3d volume for verification for each proposal\n- finally aggregate to study level\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fe0ee27a00d4956943ab75d52193c12df%2FSelection_073.png?generation=1720850421846598&alt=media)\n\n[plan-B] exit plan\n- detect all level (x,y) (we need external data for this because we need x,y for all level in train)\n- then crop 3d volume for each level to predict condition\n- finally aggregate to study level\n\n[plan-C] exit plan\n- make heatmap level(we can use kaggle data only or/and external data in train)\n- we do not predict x,y. but instead, the heatmap are additional input channel to a final image/volume classifier. \n- final image/volume (or image seq) classifier. ", "2919438": "how to read MRI:\n\nsearch keyword: Lumbar Spinal Stenosis, Lumbar MRI, beginner\nhttps://www.youtube.com/watch?v=feHB0mGnpBs\nhttps://www.youtube.com/watch?v=FkxvTsfDQ0Y\nhttps://www.youtube.com/watch?v=qYSG2KXLNyU\nhttps://www.youtube.com/watch?v=A9InC8ppa9k\nhttps://www.youtube.com/watch?v=ZtehC7q_DWc\nhttps://www.youtube.com/watch?v=qivyCxaf2IY\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F2cbd3dfc8a86041d17958612129e4b56%2FSelection_069.png?generation=1720822165743461&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fc36bfbc36e79197078b6427e2c0ada19%2FSelection_068.png?generation=1720819461134224&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F5b16d86437040adecd47457e436e670c%2FSelection_070.png?generation=1720822363328885&alt=media)\n\nmedical knowledge:\nSpine Anatomy | Know Your Spine\nhttps://www.youtube.com/watch?v=gUG_zbKqlaU\nhttps://www.youtube.com/watch?v=O-zR_YRJNhM\nhttps://www.youtube.com/watch?v=YtAk1GxtlUk\nhttps://www.youtube.com/watch?v=aQyhj9V8UxE", "2922234": "good news. initial resnet18 proposal network is successful.\ncode to release later ...\n\nleft to right: input, truth, predict\n\n ![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F03c12bffb744fc54a20fa8c0e18f349d%2FSelection_094.png?generation=1721035171133361&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F178b364f42f66b3b46c018ccd171093b%2FSelection_092.png?generation=1721035187259184&alt=media)", "2927495": "This discussion post is a gemstone. Definitely going to spend quality time on it. ", "2923507": "yet another out of box solution:\nhttps://github.com/rwindsor1/SpineNet\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F4869f5d1fbb881b41e7d3024e1e1dea6%2FSelection_105.png?generation=1721089841498777&alt=media)", "2918318": "how to use point annotation:\n[5] A New Window Loss Function for Bone Fracture Detection and Localization in X-ray Images with Point-based Annotation\nhttps://www.semanticscholar.org/reader/a5ac9aca8ea61b0126729a602e91b3886e345887\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fea23e16a6a9a3a5d2e438b3c7865fae0%2FSelection_059.png?generation=1720765762146035&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fae787ea6e8b33eca02b0e1b703af1b5f%2FSelection_060.png?generation=1720765776587799&alt=media)", "2919954": "commerical product: RadiSpine\nhttps://www.youtube.com/watch?v=F5e9uYc0xHs\nhttps://www.radirad.com/radispine-features", "2919377": "other papers that i have no time to read in details but are very much related. Maybe someone could get chatgpt to read them.\n\n- Deep learning-based high-accuracy quantitation for lumbar intervertebral disc degeneration from MRI\nhttps://www.nature.com/articles/s41467-022-28387-5\ncode: https://github.com/no-saint-no-angel/BianqueNet\n\n- Deep learning-based high-accuracy detection for lumbar and cervical degenerative disease on T2-weighted MR images\nhttps://www.researchgate.net/publication/369413726_Deep_learning-based_high-accuracy_detection_for_lumbar_and_cervical_degenerative_disease_on_T2-weighted_MR_images\n(3D ResNet18 and transformer(multi-view fusion))\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F983c6a251387a6b392b0a95ab318f1a2%2FSelection_064.png?generation=1720814884240351&alt=media)\n\n- Context-Aware Transformers For Spinal Cancer Detection and Radiological Grading\nhttps://arxiv.org/pdf/2206.13173\n(2D ResNet18 + seq + patch and transformer(multi-view fusion))\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ff2e701a22f1f5e9ca4f46c94f019cdb4%2FSelection_067.png?generation=1720815606424594&alt=media)\n\n\n- A deep learning model for detection of cervical spinal cord compression in MRI scans\nhttps://www.nature.com/articles/s41598-021-89848-3\n(resnet50 CNN, only axial images)\n", "2918304": "[1] SpineOne: A One-Stage Detection Framework for Degenerative Discs and Vertebrae\nhttps://arxiv.org/pdf/2110.15082\n\nthis work uses keypoint:\n\"we propose a one-stage detection framework termed SpineOne to simultaneously localize and classify degenerative discs and vertebrae from MRI slices. SpineOne is built upon the following three key techniques: 1) a new design of the keypoint heatmap to facilitate simultaneous keypoint localization and classification; 2) ...\"\n\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fd2c855964ef9e449b43ea530514ff706%2FSelection_058.png?generation=1720765268300160&alt=media)", "2918272": "paper[2] is an object detection based method using faster-rcnn\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fa38ee434344d1b4114172ce13e10e265%2FSelection_056.png?generation=1720763809013090&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fe2f2a409f0ac14b36483e309d7b356b9%2FSelection_057.png?generation=1720763824492488&alt=media)\n", "3003238": "Both research papers are having great insights. Thank you for sharing.\nI have few questions about it, please help me to understand following.\n1. As model is works on 2 stages, ROI detection and classification. How and what to provide I/p( like which instance of mri image) to Roi detection. As for orientation sagital there multiple slices and only few slices has clear visibility of disc?\n\n2. As in paper one after detection of verbetrals cord and disc, each disc images get detected as forms 3d volume and their respective axial orientation images for sub-articular stenosis. how each disc level relation is getting establish for axial and sagital?", "2928181": "Regarding the multi-input, mult-task, multi-class model described in paper [1], how does one structure the target labels such that we have both condition at disc level such as left_foraminal_l1/l2 (multi-task) and the severity classes (multi-class)? ", "2923531": "Great! Keep going! Looking forward for more coming today", "2922525": "Thank you hengck so much for providing so many useful papers. I always admire ability to find amazing papers. May I ask where did you find them? do you have a specific place to go to and/or method for finding those related papers? ", "2922469": "I have been trying to replicate Deep SPINE research work. It has good intuition for addressing multi-input multi-task and multi-class classification.\nIf, anyone else is also trying to work on this, kindly get in touch. Looking for a teammate! \n", "2918811": "awesome papers!! gonna check them out", "3013232": "Very helpful! Thanks", "2920545": "lot of valuable info. Thanks." } }