{ "id": 350760, "title": "Some helping tool and code (e.g. align segmentation orientation)", "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/350760", "author_name": "hengck23", "post_date": "2022-09-07T01:02:19.664000", "votes": 26, "comment_count": 9, "views": 0, "content": "

it is helpful (and important) to visualise the annotation and results.
\nthere are many free dicom viewer.
\nhere, i am using ITK-SNAP

\n

code to realign segmentation orientation

\n

update!!! you probably need to cast to .astype(np.uint8) for seg_data below

\n
import nibabel as nib\n\n#https://stackoverflow.com/questions/66231763/how-to-check-nifti-image-is-in-the-right-orientation-position-with-python\ndef set_orientation_like(seg, nii):\n    nii_xyz = nib.aff2axcodes(nii.affine)\n    seg_data = seg.get_fdata().astype(np.uint8)  ##updated!!!\n    seg_xyz  = nib.aff2axcodes(seg.affine)\n\n    if nii_xyz[0] != seg_xyz[0]: #'L':\n        seg_data = np.flip(seg_data, axis=0)\n    if nii_xyz[1] != seg_xyz[1]: # 'P':\n        seg_data = np.flip(seg_data, axis=1)\n    if nii_xyz[2] != seg_xyz[2]: # 'S':\n        seg_data = np.flip(seg_data, axis=2)\n    fixed = nib.Nifti1Image(seg_data, nii.affine, nii.header)\n    return fixed\n
\n

\"Selection-161\"

", "messages": [ { "id": 1929305, "postDate": "2022-09-07T01:02:19.663Z", "content": "

it is helpful (and important) to visualise the annotation and results.
\nthere are many free dicom viewer.
\nhere, i am using ITK-SNAP

\n

code to realign segmentation orientation

\n

update!!! you probably need to cast to .astype(np.uint8) for seg_data below

\n
import nibabel as nib\n\n#https://stackoverflow.com/questions/66231763/how-to-check-nifti-image-is-in-the-right-orientation-position-with-python\ndef set_orientation_like(seg, nii):\n    nii_xyz = nib.aff2axcodes(nii.affine)\n    seg_data = seg.get_fdata().astype(np.uint8)  ##updated!!!\n    seg_xyz  = nib.aff2axcodes(seg.affine)\n\n    if nii_xyz[0] != seg_xyz[0]: #'L':\n        seg_data = np.flip(seg_data, axis=0)\n    if nii_xyz[1] != seg_xyz[1]: # 'P':\n        seg_data = np.flip(seg_data, axis=1)\n    if nii_xyz[2] != seg_xyz[2]: # 'S':\n        seg_data = np.flip(seg_data, axis=2)\n    fixed = nib.Nifti1Image(seg_data, nii.affine, nii.header)\n    return fixed\n
\n

\"Selection-161\"

", "rawMarkdown": "it is helpful (and important) to visualise the annotation and results. \nthere are many free dicom viewer. \nhere, i am using ITK-SNAP\n\ncode to realign segmentation orientation\n\nupdate!!! you probably need to cast to .astype(np.uint8) for seg\\_data below\n```\nimport nibabel as nib\n\n#https://stackoverflow.com/questions/66231763/how-to-check-nifti-image-is-in-the-right-orientation-position-with-python\ndef set_orientation_like(seg, nii):\n nii_xyz = nib.aff2axcodes(nii.affine)\n seg_data = seg.get_fdata().astype(np.uint8) ##updated!!!\n seg_xyz = nib.aff2axcodes(seg.affine)\n\n if nii_xyz[0] != seg_xyz[0]: #'L':\n seg_data = np.flip(seg_data, axis=0)\n if nii_xyz[1] != seg_xyz[1]: # 'P':\n seg_data = np.flip(seg_data, axis=1)\n if nii_xyz[2] != seg_xyz[2]: # 'S':\n seg_data = np.flip(seg_data, axis=2)\n fixed = nib.Nifti1Image(seg_data, nii.affine, nii.header)\n return fixed\n\n\n\n```\n\n\n\n\"Selection-161\"", "votes": 26 }, { "id": 1929602, "postDate": "2022-09-07T08:00:33.280Z", "content": "

this is how fracture looks like
\n\"Selection-169\"

\n

i worked with detecting and localising fracture in x-rays images before. Last time i used about 60k image-labelled xray images and the model is able to localize fracture up to accuracy of about 90%. This can be revealed by the CAM heatmap. I used resnet50d on 2048x2048 xray images.

\n

I believed the results for spine CT would be good if we can get many,many train samples. But again getting train samples itself could be a problem in medical applications

\n
\n

at least for my previous work on xray, fracture for one type bones is useful for detection on other bones.

", "rawMarkdown": "this is how fracture looks like\n\"Selection-169\"\n\n\ni worked with detecting and localising fracture in x-rays images before. Last time i used about 60k image-labelled xray images and the model is able to localize fracture up to accuracy of about 90%. This can be revealed by the CAM heatmap. I used resnet50d on 2048x2048 xray images.\n\nI believed the results for spine CT would be good if we can get many,many train samples. But again getting train samples itself could be a problem in medical applications\n\n---\n\nat least for my previous work on xray, fracture for one type bones is useful for detection on other bones.\n\n", "votes": 1, "replies": [ { "id": 1929640, "postDate": "2022-09-07T08:31:36.533Z", "content": "

On which dataset did you get 90% accuracy for bone fracture detection ?

", "rawMarkdown": "On which dataset did you get 90% accuracy for bone fracture detection ?" }, { "id": 1929656, "postDate": "2022-09-07T08:57:23.760Z", "content": "

private hospital dataset from my client

", "rawMarkdown": "private hospital dataset from my client" }, { "id": 1929666, "postDate": "2022-09-07T09:04:05.117Z", "content": "

there has been discussion on how to use aux label like segmentation and bounding box, it should be done like this

\n

\"Selection-170\"

\n

basically aux label is used to constraint your activation map in the last layer before GAP classifier
\nhttps://www.nature.com/articles/s41467-021-21311-3.pdf

\n

then you can use human point click annotation, etc to correct pseudo aux label to create more aux label.
\nthis is what i done in my xray bones detection project

", "rawMarkdown": "there has been discussion on how to use aux label like segmentation and bounding box, it should be done like this\n\n\"Selection-170\"\n\nbasically aux label is used to constraint your activation map in the last layer before GAP classifier\nhttps://www.nature.com/articles/s41467-021-21311-3.pdf\n\nthen you can use human point click annotation, etc to correct pseudo aux label to create more aux label.\nthis is what i done in my xray bones detection project" }, { "id": 1929688, "postDate": "2022-09-07T09:23:51.990Z", "content": "

Thanks very much for sharing!

", "rawMarkdown": "Thanks very much for sharing!\n" } ] }, { "id": 1929343, "postDate": "2022-09-07T02:05:44.517Z", "content": "

3d volume rendering using 3d slicer
\nhttps://www.slicer.org/

\n

\"Selection-163\"

\n

helps to visualize the slicing

\n

\"Selection-164\"

\n

3D Slicer Tutorial - Spine Model
\nhttps://www.youtube.com/watch?v=DxHp7k9oKE8

\n

3D Slicer Tutorial: How to Segment a Lumbar Vertebrae
\nhttps://www.youtube.com/watch?v=GGgP89uTOLo
\nhttps://www.youtube.com/watch?v=Uht6Fwtr9hE&t=535s

\n

Region Growing Segmentation with 3DSlicer
\nhttps://www.youtube.com/watch?v=YGbiFh0r3EA

", "rawMarkdown": "3d volume rendering using 3d slicer \nhttps://www.slicer.org/\n\n\"Selection-163\"\n\nhelps to visualize the slicing\n\n\"Selection-164\"\n\n3D Slicer Tutorial - Spine Model\nhttps://www.youtube.com/watch?v=DxHp7k9oKE8\n\n3D Slicer Tutorial: How to Segment a Lumbar Vertebrae\nhttps://www.youtube.com/watch?v=GGgP89uTOLo\nhttps://www.youtube.com/watch?v=Uht6Fwtr9hE&t=535s\n\nRegion Growing Segmentation with 3DSlicer\nhttps://www.youtube.com/watch?v=YGbiFh0r3EA", "votes": 1, "replies": [ { "id": 1929673, "postDate": "2022-09-07T09:10:21.933Z", "content": "

for fanciful rendering

\n

https://github.com/tommybazar/TBRaymarchProject
\n\"Allows volumetric rendering of DICOM and .MHD data with Unreal Engine.\"

", "rawMarkdown": "for fanciful rendering\n\nhttps://github.com/tommybazar/TBRaymarchProject \n\"Allows volumetric rendering of DICOM and .MHD data with Unreal Engine.\" " } ] }, { "id": 1929306, "postDate": "2022-09-07T01:06:04.997Z", "content": "

code to convert dicom images to nii

\n
import SimpleITK as sitk\n\ndf = pd.read_csv(data_dir+'/train.csv')#, index_col='cell_id')\nfor i,d in df.iterrows():\n    print(i, d.StudyInstanceUID)\n    study_instance = d.StudyInstanceUID #   '1.2.826.0.1.3680043.780'\n\n    series_id = sitk.ImageSeriesReader.GetGDCMSeriesIDs(directory=dicom_dir+'/%s'%study_instance)\n    series_file  = sitk.ImageSeriesReader.GetGDCMSeriesFileNames(\n        dicom_dir+'/%s'%study_instance, series_id[0]\n    )\n    series_reader = sitk.ImageSeriesReader()\n    series_reader.SetFileNames(series_file)\n    series_reader.LoadPrivateTagsOn()\n    image = series_reader.Execute()\n\n    print(series_id)\n    print(str(series_file).replace(',',',\\n'))\n    print(image)\n\n    if 0:\n        npy = sitk.GetArrayFromImage(image)\n        npy = npy-npy.min()\n        npy = npy/npy.max()\n\n        for t in range(len(npy)):\n            image_show('npy', npy[t])\n            cv2.waitKey(0)\n\n    nii_file = nii_dir+'/%s.nii.gz'%study_instance\n    sitk.WriteImage(\n        image=image, fileName=nii_file, useCompression=True\n    )\n
", "rawMarkdown": "code to convert dicom images to nii\n\n```\nimport SimpleITK as sitk\n\ndf = pd.read_csv(data_dir+'/train.csv')#, index_col='cell_id')\nfor i,d in df.iterrows():\n print(i, d.StudyInstanceUID)\n study_instance = d.StudyInstanceUID # '1.2.826.0.1.3680043.780'\n\n series_id = sitk.ImageSeriesReader.GetGDCMSeriesIDs(directory=dicom_dir+'/%s'%study_instance)\n series_file = sitk.ImageSeriesReader.GetGDCMSeriesFileNames(\n dicom_dir+'/%s'%study_instance, series_id[0]\n )\n series_reader = sitk.ImageSeriesReader()\n series_reader.SetFileNames(series_file)\n series_reader.LoadPrivateTagsOn()\n image = series_reader.Execute()\n\n print(series_id)\n print(str(series_file).replace(',',',\\n'))\n print(image)\n\n if 0:\n npy = sitk.GetArrayFromImage(image)\n npy = npy-npy.min()\n npy = npy/npy.max()\n\n for t in range(len(npy)):\n image_show('npy', npy[t])\n cv2.waitKey(0)\n\n nii_file = nii_dir+'/%s.nii.gz'%study_instance\n sitk.WriteImage(\n image=image, fileName=nii_file, useCompression=True\n )\n\n\n```", "votes": 1 }, { "id": 1929322, "postDate": "2022-09-07T01:36:54.417Z", "content": "

you can also use itk-snap to view the 3d class activation map (CAM)
\nhow to overlay image: https://www.youtube.com/watch?v=xX5HpT67ico

\n

\"Selection-162\"

\n

in this case my model has make the decision based on noise.
\nthis method is good to debug your algorithm.

", "rawMarkdown": "you can also use itk-snap to view the 3d class activation map (CAM)\nhow to overlay image: https://www.youtube.com/watch?v=xX5HpT67ico\n\n\"Selection-162\"\n\nin this case my model has make the decision based on noise. \nthis method is good to debug your algorithm.", "votes": 2 } ], "comments": [ { "id": 1929602, "author_name": "hengck23", "author_url": "", "post_date": "2022-09-07T08:00:33.280000", "content": "

this is how fracture looks like
\n\"Selection-169\"

\n

i worked with detecting and localising fracture in x-rays images before. Last time i used about 60k image-labelled xray images and the model is able to localize fracture up to accuracy of about 90%. This can be revealed by the CAM heatmap. I used resnet50d on 2048x2048 xray images.

\n

I believed the results for spine CT would be good if we can get many,many train samples. But again getting train samples itself could be a problem in medical applications

\n
\n

at least for my previous work on xray, fracture for one type bones is useful for detection on other bones.

", "votes": 1, "replies": [ { "id": 1929640, "author_name": "Optimo", "author_url": "", "post_date": "2022-09-07T08:31:36.533000", "content": "

On which dataset did you get 90% accuracy for bone fracture detection ?

", "votes": 0, "replies": [] }, { "id": 1929656, "author_name": "hengck23", "author_url": "", "post_date": "2022-09-07T08:57:23.760000", "content": "

private hospital dataset from my client

", "votes": 0, "replies": [] }, { "id": 1929666, "author_name": "hengck23", "author_url": "", "post_date": "2022-09-07T09:04:05.117000", "content": "

there has been discussion on how to use aux label like segmentation and bounding box, it should be done like this

\n

\"Selection-170\"

\n

basically aux label is used to constraint your activation map in the last layer before GAP classifier
\nhttps://www.nature.com/articles/s41467-021-21311-3.pdf

\n

then you can use human point click annotation, etc to correct pseudo aux label to create more aux label.
\nthis is what i done in my xray bones detection project

", "votes": 0, "replies": [] }, { "id": 1929688, "author_name": "Optimo", "author_url": "", "post_date": "2022-09-07T09:23:51.990000", "content": "

Thanks very much for sharing!

", "votes": 0, "replies": [] } ] }, { "id": 1929343, "author_name": "hengck23", "author_url": "", "post_date": "2022-09-07T02:05:44.517000", "content": "

3d volume rendering using 3d slicer
\nhttps://www.slicer.org/

\n

\"Selection-163\"

\n

helps to visualize the slicing

\n

\"Selection-164\"

\n

3D Slicer Tutorial - Spine Model
\nhttps://www.youtube.com/watch?v=DxHp7k9oKE8

\n

3D Slicer Tutorial: How to Segment a Lumbar Vertebrae
\nhttps://www.youtube.com/watch?v=GGgP89uTOLo
\nhttps://www.youtube.com/watch?v=Uht6Fwtr9hE&t=535s

\n

Region Growing Segmentation with 3DSlicer
\nhttps://www.youtube.com/watch?v=YGbiFh0r3EA

", "votes": 1, "replies": [ { "id": 1929673, "author_name": "hengck23", "author_url": "", "post_date": "2022-09-07T09:10:21.933000", "content": "

for fanciful rendering

\n

https://github.com/tommybazar/TBRaymarchProject
\n\"Allows volumetric rendering of DICOM and .MHD data with Unreal Engine.\"

", "votes": 0, "replies": [] } ] }, { "id": 1929306, "author_name": "hengck23", "author_url": "", "post_date": "2022-09-07T01:06:04.997000", "content": "

code to convert dicom images to nii

\n
import SimpleITK as sitk\n\ndf = pd.read_csv(data_dir+'/train.csv')#, index_col='cell_id')\nfor i,d in df.iterrows():\n    print(i, d.StudyInstanceUID)\n    study_instance = d.StudyInstanceUID #   '1.2.826.0.1.3680043.780'\n\n    series_id = sitk.ImageSeriesReader.GetGDCMSeriesIDs(directory=dicom_dir+'/%s'%study_instance)\n    series_file  = sitk.ImageSeriesReader.GetGDCMSeriesFileNames(\n        dicom_dir+'/%s'%study_instance, series_id[0]\n    )\n    series_reader = sitk.ImageSeriesReader()\n    series_reader.SetFileNames(series_file)\n    series_reader.LoadPrivateTagsOn()\n    image = series_reader.Execute()\n\n    print(series_id)\n    print(str(series_file).replace(',',',\\n'))\n    print(image)\n\n    if 0:\n        npy = sitk.GetArrayFromImage(image)\n        npy = npy-npy.min()\n        npy = npy/npy.max()\n\n        for t in range(len(npy)):\n            image_show('npy', npy[t])\n            cv2.waitKey(0)\n\n    nii_file = nii_dir+'/%s.nii.gz'%study_instance\n    sitk.WriteImage(\n        image=image, fileName=nii_file, useCompression=True\n    )\n
", "votes": 1, "replies": [] }, { "id": 1929322, "author_name": "hengck23", "author_url": "", "post_date": "2022-09-07T01:36:54.417000", "content": "

you can also use itk-snap to view the 3d class activation map (CAM)
\nhow to overlay image: https://www.youtube.com/watch?v=xX5HpT67ico

\n

\"Selection-162\"

\n

in this case my model has make the decision based on noise.
\nthis method is good to debug your algorithm.

", "votes": 2, "replies": [] } ], "raw_markdown_by_id": { "1929305": "it is helpful (and important) to visualise the annotation and results. \nthere are many free dicom viewer. \nhere, i am using ITK-SNAP\n\ncode to realign segmentation orientation\n\nupdate!!! you probably need to cast to .astype(np.uint8) for seg\\_data below\n```\nimport nibabel as nib\n\n#https://stackoverflow.com/questions/66231763/how-to-check-nifti-image-is-in-the-right-orientation-position-with-python\ndef set_orientation_like(seg, nii):\n nii_xyz = nib.aff2axcodes(nii.affine)\n seg_data = seg.get_fdata().astype(np.uint8) ##updated!!!\n seg_xyz = nib.aff2axcodes(seg.affine)\n\n if nii_xyz[0] != seg_xyz[0]: #'L':\n seg_data = np.flip(seg_data, axis=0)\n if nii_xyz[1] != seg_xyz[1]: # 'P':\n seg_data = np.flip(seg_data, axis=1)\n if nii_xyz[2] != seg_xyz[2]: # 'S':\n seg_data = np.flip(seg_data, axis=2)\n fixed = nib.Nifti1Image(seg_data, nii.affine, nii.header)\n return fixed\n\n\n\n```\n\n\n\n\"Selection-161\"", "1929602": "this is how fracture looks like\n\"Selection-169\"\n\n\ni worked with detecting and localising fracture in x-rays images before. Last time i used about 60k image-labelled xray images and the model is able to localize fracture up to accuracy of about 90%. This can be revealed by the CAM heatmap. I used resnet50d on 2048x2048 xray images.\n\nI believed the results for spine CT would be good if we can get many,many train samples. But again getting train samples itself could be a problem in medical applications\n\n---\n\nat least for my previous work on xray, fracture for one type bones is useful for detection on other bones.\n\n", "1929343": "3d volume rendering using 3d slicer \nhttps://www.slicer.org/\n\n\"Selection-163\"\n\nhelps to visualize the slicing\n\n\"Selection-164\"\n\n3D Slicer Tutorial - Spine Model\nhttps://www.youtube.com/watch?v=DxHp7k9oKE8\n\n3D Slicer Tutorial: How to Segment a Lumbar Vertebrae\nhttps://www.youtube.com/watch?v=GGgP89uTOLo\nhttps://www.youtube.com/watch?v=Uht6Fwtr9hE&t=535s\n\nRegion Growing Segmentation with 3DSlicer\nhttps://www.youtube.com/watch?v=YGbiFh0r3EA", "1929306": "code to convert dicom images to nii\n\n```\nimport SimpleITK as sitk\n\ndf = pd.read_csv(data_dir+'/train.csv')#, index_col='cell_id')\nfor i,d in df.iterrows():\n print(i, d.StudyInstanceUID)\n study_instance = d.StudyInstanceUID # '1.2.826.0.1.3680043.780'\n\n series_id = sitk.ImageSeriesReader.GetGDCMSeriesIDs(directory=dicom_dir+'/%s'%study_instance)\n series_file = sitk.ImageSeriesReader.GetGDCMSeriesFileNames(\n dicom_dir+'/%s'%study_instance, series_id[0]\n )\n series_reader = sitk.ImageSeriesReader()\n series_reader.SetFileNames(series_file)\n series_reader.LoadPrivateTagsOn()\n image = series_reader.Execute()\n\n print(series_id)\n print(str(series_file).replace(',',',\\n'))\n print(image)\n\n if 0:\n npy = sitk.GetArrayFromImage(image)\n npy = npy-npy.min()\n npy = npy/npy.max()\n\n for t in range(len(npy)):\n image_show('npy', npy[t])\n cv2.waitKey(0)\n\n nii_file = nii_dir+'/%s.nii.gz'%study_instance\n sitk.WriteImage(\n image=image, fileName=nii_file, useCompression=True\n )\n\n\n```", "1929322": "you can also use itk-snap to view the 3d class activation map (CAM)\nhow to overlay image: https://www.youtube.com/watch?v=xX5HpT67ico\n\n\"Selection-162\"\n\nin this case my model has make the decision based on noise. \nthis method is good to debug your algorithm." } }