![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F856355%2F6e4d08b10ead7a1c9b3904f1d83bd626%2FWechatIMG1025.jpg?generation=1728957811388315&alt=media\")
\nTeam post-submission notebook can be found at:
\nhttps://www.kaggle.com/code/hengck23/post-lhw-v24-ensemble-add-heng
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F856355%2F07bfb7f227ec83b40bc3b7a92c5f7092%2FWechatIMG1026.jpg?generation=1728957822586591&alt=media\")
First, we would like to express our gratitude to Kaggle and RSNA for hosting this great competition.
\nOur solution is an ensemble of 1 stage and 2 stage.
I'm very grateful to my teammates, I learned a lot from them.
2 stage methods, components:
\nkeypoint regression (both 2d and 3d)
Sagittal:
\nAxial:
\nUsed backbones, which is searched by ITK:
\nfuse 3 view on feature space
\nThe keypoint models used are keypoint-sag-2d and keypoint-axial-3d.
\nFor the sagittal plane, we sample 10 slices for each series, cropping the center 5 slices on T2, and cropping 5 slices from the left and 5 slices from the right on T1, for each level.
\nFor the axial plane, we sample 5 slices for each level based on the Z-coordinate of the point.
\nThe final input size is: x -> (bs, 5_cond, 5_level, 5, crop_h, crop_w).
\nThe model architecture is as follows:
\ny, axial_embs = .axial_model(axial_x) \nys2, sag_embs = .sag_model(x)\n\nsag_embs = sag_embs.permute(, , , ).reshape(b, , -) \nembs = torch.cat((sag_embs, axial_embs), dim=-)\nys = .out_linear(embs).reshape(b, , , ) \nys = ys.permute(, , , ) \nys = ys + ys2\nThe defect of stage2_model1 is that it uses 2D keypoints on the sagittal plane and does not handle multiple series on the axial plane very well.
\nThe keypoint models used are keypoint-sag-3d and keypoint-axial-level-cls.
\nwe crop the 3D ROI image using (z_imgs, crop_h, crop_w) based on XYZ points, where z_imgs can be either 3 or 5; we found that 3 works best for my purposes.
\nFor the axial plane, there may be multiple detections, but we only used the one with the highest confidence (not necessarily the best).
stage2_model1 + stage2_model2 with many backbones, TTA: hflip for sagittal, vflip for axial
\nresults: cv: 0.373 LB: 0.35 PB: 0.40
backbones:
\nTeam submission notebook can be found at:
\nhttps://www.kaggle.com/code/hengck23/lhw-v24-ensemble-add-heng
\nTeam post-submission notebook can be found at:
\nhttps://www.kaggle.com/code/hengck23/post-lhw-v24-ensemble-add-heng
\n
Infer Notebook:
\nFirst, we would like to express our gratitude to Kaggle and RSNA for hosting this great competition.
\nOur solution is an ensemble of 1 stage and 2 stage.
I'm very grateful to my teammates, I learned a lot from them.
2 stage methods, components:
\nkeypoint regression (both 2d and 3d)
Sagittal:
\nAxial:
\nUsed backbones, which is searched by ITK:
\nfuse 3 view on feature space
\nThe keypoint models used are keypoint-sag-2d and keypoint-axial-3d.
\nFor the sagittal plane, we sample 10 slices for each series, cropping the center 5 slices on T2, and cropping 5 slices from the left and 5 slices from the right on T1, for each level.
\nFor the axial plane, we sample 5 slices for each level based on the Z-coordinate of the point.
\nThe final input size is: x -> (bs, 5_cond, 5_level, 5, crop_h, crop_w).
\nThe model architecture is as follows:
\ny, axial_embs = .axial_model(axial_x) \nys2, sag_embs = .sag_model(x)\n\nsag_embs = sag_embs.permute(, , , ).reshape(b, , -) \nembs = torch.cat((sag_embs, axial_embs), dim=-)\nys = .out_linear(embs).reshape(b, , , ) \nys = ys.permute(, , , ) \nys = ys + ys2\nThe defect of stage2_model1 is that it uses 2D keypoints on the sagittal plane and does not handle multiple series on the axial plane very well.
\nThe keypoint models used are keypoint-sag-3d and keypoint-axial-level-cls.
\nwe crop the 3D ROI image using (z_imgs, crop_h, crop_w) based on XYZ points, where z_imgs can be either 3 or 5; we found that 3 works best for my purposes.
\nFor the axial plane, there may be multiple detections, but we only used the one with the highest confidence (not necessarily the best).
stage2_model1 + stage2_model2 with many backbones, TTA: hflip for sagittal, vflip for axial
\nresults: cv: 0.373 LB: 0.35 PB: 0.40
backbones:
\nTeam submission notebook can be found at:
\nhttps://www.kaggle.com/code/hengck23/lhw-v24-ensemble-add-heng
\nTeam post-submission notebook can be found at:
\nhttps://www.kaggle.com/code/hengck23/post-lhw-v24-ensemble-add-heng
\n
Infer Notebook:
\nCongratulations on your seventh place. Do you have a specific notebook
", "rawMarkdown": "Congratulations on your seventh place. Do you have a specific notebook", "votes": null }, { "id": "3014773", "postDate": "10/11/2024 15:20:01", "content": "Looking forward to your code, thanks.
", "rawMarkdown": "Looking forward to your code, thanks.", "votes": null } ], "comments": [ { "id": 3012928, "author_name": "icw1lee", "author_url": "", "post_date": "10/09/2024 14:01:38", "content": "Congratulations on your seventh place. Do you have a specific notebook
", "votes": null, "replies": [] }, { "id": 3014773, "author_name": "nextnlp", "author_url": "", "post_date": "10/11/2024 15:20:01", "content": "Looking forward to your code, thanks.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3012554": "# Summary\nFirst, we would like to express our gratitude to Kaggle and RSNA for hosting this great competition.