{ "id": 512048, "title": "Suspicious wrong coordinates labels", "url": "/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/512048", "author_name": "Hajimi Nanbeilvdou", "post_date": "2024-06-13T08:51:50.013000", "votes": 18, "comment_count": 10, "views": 0, "content": "

Hi, everybody. I found some coordinates may be incorrectly annotated on images of stdudy id 4279881930 series id 3657125167 instance number 14, 15 and 16. Compared with images ahead like instance 7 and 8, the coordinate level seems get wrongly labled higher by one. (e.g. it should be l3/l4 but it is labeled as l2/l3)
\nI am not an expert and hope to hear any opinions.
\n\"4279881930_3657125167_15

", "messages": [ { "id": 2869764, "postDate": "2024-06-13T08:51:50.013Z", "content": "

Hi, everybody. I found some coordinates may be incorrectly annotated on images of stdudy id 4279881930 series id 3657125167 instance number 14, 15 and 16. Compared with images ahead like instance 7 and 8, the coordinate level seems get wrongly labled higher by one. (e.g. it should be l3/l4 but it is labeled as l2/l3)
\nI am not an expert and hope to hear any opinions.
\n\"4279881930_3657125167_15

", "rawMarkdown": "Hi, everybody. I found some coordinates may be incorrectly annotated on images of stdudy id 4279881930 series id 3657125167 instance number 14, 15 and 16. Compared with images ahead like instance 7 and 8, the coordinate level seems get wrongly labled higher by one. (e.g. it should be l3/l4 but it is labeled as l2/l3)\nI am not an expert and hope to hear any opinions.\n![4279881930_3657125167_15 ](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5567488%2Fc5784a6da9d135c92d1f9188af9c4f72%2F4279881930_3657125167_15.png?generation=1718268673559803&alt=media)", "votes": 18 }, { "id": 2870939, "postDate": "2024-06-13T22:42:17.043Z", "content": "

", "rawMarkdown": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6259210%2Fef9522446b7fc46900643a4861319803%2F5examples.png?generation=1718329552852857&alt=media)", "votes": 5, "replies": [ { "id": 2871177, "postDate": "2024-06-14T05:25:35.933Z", "content": "

Thanks for pointing this out. In addition, there seems to be a (5, 5) error in some of the slices. I have listed them here:

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
study_idseries_idinstance_numberconditionlevelxy
382814208803611569Spinal Canal StenosisL5/S155
286903519192191720513Spinal Canal StenosisL5/S15.000015.00001
66562726322314716339Spinal Canal StenosisL1/L255
14387605437377538159Spinal Canal StenosisL1/L25.000044.99999
151045189714888575509Spinal Canal StenosisL4/L554.99998
188097048037369415258Spinal Canal StenosisL5/S155
1901348744149027245611Spinal Canal StenosisL3/L44.999954.99997
215146750730867193298Spinal Canal StenosisL3/L44.058825
2316015842148519329913Spinal Canal StenosisL1/L25.000035.00001
2444340715352140919810Spinal Canal StenosisL5/S15.000034.99999
290502590481638137811Spinal Canal StenosisL1/L24.999985
\n

These would be equivalent to the 1st (38281420, 880361156), 2nd (286903519, 1921917205) and last slice (2905025904, 816381378) above.

", "rawMarkdown": "Thanks for pointing this out. In addition, there seems to be a (5, 5) error in some of the slices. I have listed them here:\n| study_id | series_id | instance_number | condition | level | x | y |\n|-----------:|------------:|------------------:|:----------------------|:--------|--------:|--------:|\n| 38281420 | 880361156 | 9 | Spinal Canal Stenosis | L5/S1 | 5 | 5 |\n| 286903519 | 1921917205 | 13 | Spinal Canal Stenosis | L5/S1 | 5.00001 | 5.00001 |\n| 665627263 | 2231471633 | 9 | Spinal Canal Stenosis | L1/L2 | 5 | 5 |\n| 1438760543 | 737753815 | 9 | Spinal Canal Stenosis | L1/L2 | 5.00004 | 4.99999 |\n| 1510451897 | 1488857550 | 9 | Spinal Canal Stenosis | L4/L5 | 5 | 4.99998 |\n| 1880970480 | 3736941525 | 8 | Spinal Canal Stenosis | L5/S1 | 5 | 5 |\n| 1901348744 | 1490272456 | 11 | Spinal Canal Stenosis | L3/L4 | 4.99995 | 4.99997 |\n| 2151467507 | 3086719329 | 8 | Spinal Canal Stenosis | L3/L4 | 4.05882 | 5 |\n| 2316015842 | 1485193299 | 13 | Spinal Canal Stenosis | L1/L2 | 5.00003 | 5.00001 |\n| 2444340715 | 3521409198 | 10 | Spinal Canal Stenosis | L5/S1 | 5.00003 | 4.99999 |\n| 2905025904 | 816381378 | 11 | Spinal Canal Stenosis | L1/L2 | 4.99998 | 5 |\n\nThese would be equivalent to the 1st (38281420, 880361156), 2nd (286903519, 1921917205) and last slice (2905025904, 816381378) above.", "votes": 6 }, { "id": 2874163, "postDate": "2024-06-16T04:59:12.077Z", "content": "

Thanks a lot. For pointing it out in a clear way.

", "rawMarkdown": "Thanks a lot. For pointing it out in a clear way." } ] }, { "id": 2870341, "postDate": "2024-06-13T14:23:18.957Z", "content": "

I am not an expert by any means, but a side-by-side comparison of the slices reveal that at least the adjacent slices are consistent:
\n
\nHere is the code to create this:

\n
 pathlib  Path\n\n matplotlib.pyplot  plt\n pandas  pd\n pydicom\n\nstudy_id, series_id = , \nfig, axes = plt.subplots(, , figsize=(, ))\n\ninstance_numbers = [, , ]\n instance_number, ax  (instance_numbers, axes[]):\n    dcm_path = train_images_path / (study_id) / (series_id) / \n     dcm_path.exists()\n    dcm = pydicom.dcmread(dcm_path)\n     (dcm.InstanceNumber) == instance_number \n\n    ax.imshow(dcm.pixel_array, )\n\n    conditions = (\n        (df_train_label.study_id == study_id) & (df_train_label.series_id == series_id) & (df_train_label.instance_number == instance_number)\n    )\n    sub_lbls = df_train_label[conditions]\n    sub_lbls.plot(, , , ax=ax, color=)\n    ax.set_title( + .join(sub_lbls.level.values.tolist()))\n\ninstance_numbers = [, , ]\n instance_number, ax  (instance_numbers, axes[]):\n    dcm_path = train_images_path / (study_id) / (series_id) / \n     dcm_path.exists()\n    dcm = pydicom.dcmread(dcm_path)\n     (dcm.InstanceNumber) == instance_number \n\n    ax.imshow(dcm.pixel_array, )\n\n    conditions = (\n        (df_train_label.study_id == study_id) & (df_train_label.series_id == series_id) & (df_train_label.instance_number == instance_number)\n    )\n    sub_lbls = df_train_label[conditions]\n    sub_lbls.plot(, , , ax=ax, color=)\n    ax.set_title( + .join(sub_lbls.level.values.tolist()))\nplt.show()\n
", "rawMarkdown": "I am not an expert by any means, but a side-by-side comparison of the slices reveal that at least the adjacent slices are consistent:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1303569%2Fa298e0c69d9df27fe40d7983ebb52bb2%2Fside-by-side_slices.png?generation=1718288436785505&alt=media)\nHere is the code to create this:\n```python\nfrom pathlib import Path\n\nimport matplotlib.pyplot as plt\nimport pandas as pd\nimport pydicom\n\nstudy_id, series_id = 4279881930, 3657125167\nfig, axes = plt.subplots(2, 3, figsize=(20, 10))\n\ninstance_numbers = [6, 7, 8]\nfor instance_number, ax in zip(instance_numbers, axes[0]):\n dcm_path = train_images_path / str(study_id) / str(series_id) / f\"{instance_number}.dcm\"\n assert dcm_path.exists()\n dcm = pydicom.dcmread(dcm_path)\n assert int(dcm.InstanceNumber) == instance_number # Double check\n\n ax.imshow(dcm.pixel_array, \"gray\")\n\n conditions = (\n (df_train_label.study_id == study_id) & (df_train_label.series_id == series_id) & (df_train_label.instance_number == instance_number)\n )\n sub_lbls = df_train_label[conditions]\n sub_lbls.plot(\"x\", \"y\", \"scatter\", ax=ax, color=\"red\")\n ax.set_title(f\"{instance_number}:\" + \", \".join(sub_lbls.level.values.tolist()))\n \ninstance_numbers = [14, 15, 16]\nfor instance_number, ax in zip(instance_numbers, axes[1]):\n dcm_path = train_images_path / str(study_id) / str(series_id) / f\"{instance_number}.dcm\"\n assert dcm_path.exists()\n dcm = pydicom.dcmread(dcm_path)\n assert int(dcm.InstanceNumber) == instance_number # Double check\n\n ax.imshow(dcm.pixel_array, \"gray\")\n\n conditions = (\n (df_train_label.study_id == study_id) & (df_train_label.series_id == series_id) & (df_train_label.instance_number == instance_number)\n )\n sub_lbls = df_train_label[conditions]\n sub_lbls.plot(\"x\", \"y\", \"scatter\", ax=ax, color=\"red\")\n ax.set_title(f\"{instance_number}:\" + \", \".join(sub_lbls.level.values.tolist()))\nplt.show()\n```", "votes": 1 }, { "id": 2874166, "postDate": "2024-06-16T05:02:17.787Z", "content": "

Thank you. And how did you manage to find it out? Were you going through each image and checking if they are labelled correctly?

", "rawMarkdown": "Thank you. And how did you manage to find it out? Were you going through each image and checking if they are labelled correctly?", "replies": [ { "id": 2874443, "postDate": "2024-06-16T09:14:25.507Z", "content": "

In my case I noticed in the examples @sergiosaharovskiy put above in comment, that 3 of the 5 images showcased points close to (5, 5). I filtered for x and y close to 5 and got the list I outlined in the reply.

", "rawMarkdown": "In my case I noticed in the examples @sergiosaharovskiy put above in comment, that 3 of the 5 images showcased points close to (5, 5). I filtered for x and y close to 5 and got the list I outlined in the reply.", "votes": 2, "replies": [ { "id": 2874476, "postDate": "2024-06-16T10:01:18.460Z", "content": "

You mean that you applied a filter for each unique image. The filter returned the images that had multiple labels in them whose coordinates were less than 5 units (pixels?) away from each other.

\n

Do you mean this?

", "rawMarkdown": "You mean that you applied a filter for each unique image. The filter returned the images that had multiple labels in them whose coordinates were less than 5 units (pixels?) away from each other.\n\nDo you mean this?" }, { "id": 2874480, "postDate": "2024-06-16T10:05:14.547Z", "content": "

Ok, now I know how you found this. Thank you

\n

The approach in my previous comment can also be very effective (in the case of sagittal views ) because correct coordinates in a single image may be at least somewhat far from each other .

", "rawMarkdown": "Ok, now I know how you found this. Thank you\n\nThe approach in my previous comment can also be very effective (in the case of sagittal views ) because correct coordinates in a single image may be at least somewhat far from each other .", "votes": 1 }, { "id": 2875381, "postDate": "2024-06-17T05:12:40.827Z", "content": "

This is the exact code for getting the above table for coordinates near (5.0, 5.0):

\n
 pathlib  Path\n\n pandas  pd\n\n\nINPUT_DIR = Path()\ntrain_images_path = INPUT_DIR / \n\n\ndf_train_main = pd.read_csv(INPUT_DIR / )\ndf_train_label = pd.read_csv(INPUT_DIR / )\n\n(df_train_label[((df_train_label.x - ) **  + (df_train_label.y - ) ** ) <= ].to_markdown(index=))\n
", "rawMarkdown": "This is the exact code for getting the above table for coordinates near (5.0, 5.0):\n```python\nfrom pathlib import Path\n\nimport pandas as pd\n\n\nINPUT_DIR = Path(\"../input/rsna-2024-lumbar-spine-degenerative-classification\")\ntrain_images_path = INPUT_DIR / \"train_images\"\n\n# read data\ndf_train_main = pd.read_csv(INPUT_DIR / 'train.csv')\ndf_train_label = pd.read_csv(INPUT_DIR / 'train_label_coordinates.csv')\n\nprint(df_train_label[((df_train_label.x - 5.0) ** 2 + (df_train_label.y - 5.0) ** 2) <= 1].to_markdown(index=False))\n```", "votes": 1 } ] } ] }, { "id": 2871964, "postDate": "2024-06-14T13:46:50.923Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 2870939, "author_name": "SSS", "author_url": "", "post_date": "2024-06-13T22:42:17.043000", "content": "

", "votes": 5, "replies": [ { "id": 2871177, "author_name": "coderRKJ", "author_url": "", "post_date": "2024-06-14T05:25:35.933000", "content": "

Thanks for pointing this out. In addition, there seems to be a (5, 5) error in some of the slices. I have listed them here:

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
study_idseries_idinstance_numberconditionlevelxy
382814208803611569Spinal Canal StenosisL5/S155
286903519192191720513Spinal Canal StenosisL5/S15.000015.00001
66562726322314716339Spinal Canal StenosisL1/L255
14387605437377538159Spinal Canal StenosisL1/L25.000044.99999
151045189714888575509Spinal Canal StenosisL4/L554.99998
188097048037369415258Spinal Canal StenosisL5/S155
1901348744149027245611Spinal Canal StenosisL3/L44.999954.99997
215146750730867193298Spinal Canal StenosisL3/L44.058825
2316015842148519329913Spinal Canal StenosisL1/L25.000035.00001
2444340715352140919810Spinal Canal StenosisL5/S15.000034.99999
290502590481638137811Spinal Canal StenosisL1/L24.999985
\n

These would be equivalent to the 1st (38281420, 880361156), 2nd (286903519, 1921917205) and last slice (2905025904, 816381378) above.

", "votes": 6, "replies": [] }, { "id": 2874163, "author_name": "Devsya ", "author_url": "", "post_date": "2024-06-16T04:59:12.077000", "content": "

Thanks a lot. For pointing it out in a clear way.

", "votes": 0, "replies": [] } ] }, { "id": 2870341, "author_name": "coderRKJ", "author_url": "", "post_date": "2024-06-13T14:23:18.957000", "content": "

I am not an expert by any means, but a side-by-side comparison of the slices reveal that at least the adjacent slices are consistent:
\n
\nHere is the code to create this:

\n
 pathlib  Path\n\n matplotlib.pyplot  plt\n pandas  pd\n pydicom\n\nstudy_id, series_id = , \nfig, axes = plt.subplots(, , figsize=(, ))\n\ninstance_numbers = [, , ]\n instance_number, ax  (instance_numbers, axes[]):\n    dcm_path = train_images_path / (study_id) / (series_id) / \n     dcm_path.exists()\n    dcm = pydicom.dcmread(dcm_path)\n     (dcm.InstanceNumber) == instance_number \n\n    ax.imshow(dcm.pixel_array, )\n\n    conditions = (\n        (df_train_label.study_id == study_id) & (df_train_label.series_id == series_id) & (df_train_label.instance_number == instance_number)\n    )\n    sub_lbls = df_train_label[conditions]\n    sub_lbls.plot(, , , ax=ax, color=)\n    ax.set_title( + .join(sub_lbls.level.values.tolist()))\n\ninstance_numbers = [, , ]\n instance_number, ax  (instance_numbers, axes[]):\n    dcm_path = train_images_path / (study_id) / (series_id) / \n     dcm_path.exists()\n    dcm = pydicom.dcmread(dcm_path)\n     (dcm.InstanceNumber) == instance_number \n\n    ax.imshow(dcm.pixel_array, )\n\n    conditions = (\n        (df_train_label.study_id == study_id) & (df_train_label.series_id == series_id) & (df_train_label.instance_number == instance_number)\n    )\n    sub_lbls = df_train_label[conditions]\n    sub_lbls.plot(, , , ax=ax, color=)\n    ax.set_title( + .join(sub_lbls.level.values.tolist()))\nplt.show()\n
", "votes": 1, "replies": [] }, { "id": 2874166, "author_name": "Devsya ", "author_url": "", "post_date": "2024-06-16T05:02:17.787000", "content": "

Thank you. And how did you manage to find it out? Were you going through each image and checking if they are labelled correctly?

", "votes": 0, "replies": [ { "id": 2874443, "author_name": "coderRKJ", "author_url": "", "post_date": "2024-06-16T09:14:25.507000", "content": "

In my case I noticed in the examples @sergiosaharovskiy put above in comment, that 3 of the 5 images showcased points close to (5, 5). I filtered for x and y close to 5 and got the list I outlined in the reply.

", "votes": 2, "replies": [ { "id": 2874476, "author_name": "Devsya ", "author_url": "", "post_date": "2024-06-16T10:01:18.460000", "content": "

You mean that you applied a filter for each unique image. The filter returned the images that had multiple labels in them whose coordinates were less than 5 units (pixels?) away from each other.

\n

Do you mean this?

", "votes": 0, "replies": [] }, { "id": 2874480, "author_name": "Devsya ", "author_url": "", "post_date": "2024-06-16T10:05:14.547000", "content": "

Ok, now I know how you found this. Thank you

\n

The approach in my previous comment can also be very effective (in the case of sagittal views ) because correct coordinates in a single image may be at least somewhat far from each other .

", "votes": 1, "replies": [] }, { "id": 2875381, "author_name": "coderRKJ", "author_url": "", "post_date": "2024-06-17T05:12:40.827000", "content": "

This is the exact code for getting the above table for coordinates near (5.0, 5.0):

\n
 pathlib  Path\n\n pandas  pd\n\n\nINPUT_DIR = Path()\ntrain_images_path = INPUT_DIR / \n\n\ndf_train_main = pd.read_csv(INPUT_DIR / )\ndf_train_label = pd.read_csv(INPUT_DIR / )\n\n(df_train_label[((df_train_label.x - ) **  + (df_train_label.y - ) ** ) <= ].to_markdown(index=))\n
", "votes": 1, "replies": [] } ] } ] }, { "id": 2871964, "author_name": "", "author_url": "", "post_date": "2024-06-14T13:46:50.923000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2869764": "Hi, everybody. I found some coordinates may be incorrectly annotated on images of stdudy id 4279881930 series id 3657125167 instance number 14, 15 and 16. Compared with images ahead like instance 7 and 8, the coordinate level seems get wrongly labled higher by one. (e.g. it should be l3/l4 but it is labeled as l2/l3)\nI am not an expert and hope to hear any opinions.\n![4279881930_3657125167_15 ](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5567488%2Fc5784a6da9d135c92d1f9188af9c4f72%2F4279881930_3657125167_15.png?generation=1718268673559803&alt=media)", "2870939": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6259210%2Fef9522446b7fc46900643a4861319803%2F5examples.png?generation=1718329552852857&alt=media)", "2870341": "I am not an expert by any means, but a side-by-side comparison of the slices reveal that at least the adjacent slices are consistent:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1303569%2Fa298e0c69d9df27fe40d7983ebb52bb2%2Fside-by-side_slices.png?generation=1718288436785505&alt=media)\nHere is the code to create this:\n```python\nfrom pathlib import Path\n\nimport matplotlib.pyplot as plt\nimport pandas as pd\nimport pydicom\n\nstudy_id, series_id = 4279881930, 3657125167\nfig, axes = plt.subplots(2, 3, figsize=(20, 10))\n\ninstance_numbers = [6, 7, 8]\nfor instance_number, ax in zip(instance_numbers, axes[0]):\n dcm_path = train_images_path / str(study_id) / str(series_id) / f\"{instance_number}.dcm\"\n assert dcm_path.exists()\n dcm = pydicom.dcmread(dcm_path)\n assert int(dcm.InstanceNumber) == instance_number # Double check\n\n ax.imshow(dcm.pixel_array, \"gray\")\n\n conditions = (\n (df_train_label.study_id == study_id) & (df_train_label.series_id == series_id) & (df_train_label.instance_number == instance_number)\n )\n sub_lbls = df_train_label[conditions]\n sub_lbls.plot(\"x\", \"y\", \"scatter\", ax=ax, color=\"red\")\n ax.set_title(f\"{instance_number}:\" + \", \".join(sub_lbls.level.values.tolist()))\n \ninstance_numbers = [14, 15, 16]\nfor instance_number, ax in zip(instance_numbers, axes[1]):\n dcm_path = train_images_path / str(study_id) / str(series_id) / f\"{instance_number}.dcm\"\n assert dcm_path.exists()\n dcm = pydicom.dcmread(dcm_path)\n assert int(dcm.InstanceNumber) == instance_number # Double check\n\n ax.imshow(dcm.pixel_array, \"gray\")\n\n conditions = (\n (df_train_label.study_id == study_id) & (df_train_label.series_id == series_id) & (df_train_label.instance_number == instance_number)\n )\n sub_lbls = df_train_label[conditions]\n sub_lbls.plot(\"x\", \"y\", \"scatter\", ax=ax, color=\"red\")\n ax.set_title(f\"{instance_number}:\" + \", \".join(sub_lbls.level.values.tolist()))\nplt.show()\n```", "2874166": "Thank you. And how did you manage to find it out? Were you going through each image and checking if they are labelled correctly?", "2871964": "" } }