![\"rsna\"](\"https://i.ibb.co/RpyFV6g/rsna.png\")
This was a really interesting competition, thanks to the hosts for organizing another cool medical challenge.
\nThanks also to my teammates @optimo and @tivfrvqhs5 for the awesome competition. David and I joined @optimo's team shortly before the team-merge deadline when he was already at LB 0.4, which gave gave us a week to improve his already great pipeline.
I will keep the write-up short but feel free to ask any question in the comments.
\n
Our overall solution is a 3 stage pipeline with additional stuff on the side, illustrated above.
\nThanks for reading, hope to see you all on the next medical imaging comp =)
", "messages": [ { "id": 2007491, "postDate": "2022-10-28T09:11:30.620Z", "content": "This was a really interesting competition, thanks to the hosts for organizing another cool medical challenge.
\nThanks also to my teammates @optimo and @tivfrvqhs5 for the awesome competition. David and I joined @optimo's team shortly before the team-merge deadline when he was already at LB 0.4, which gave gave us a week to improve his already great pipeline.
I will keep the write-up short but feel free to ask any question in the comments.
\n
Our overall solution is a 3 stage pipeline with additional stuff on the side, illustrated above.
\nThanks for reading, hope to see you all on the next medical imaging comp =)
", "rawMarkdown": "#### Intro\n\nThis was a really interesting competition, thanks to the hosts for organizing another cool medical challenge.\nThanks also to my teammates @optimo and @tivfrvqhs5 for the awesome competition. David and I joined @optimo's team shortly before the team-merge deadline when he was already at LB 0.4, which gave gave us a week to improve his already great pipeline.\n\nI will keep the write-up short but feel free to ask any question in the comments.\n\n#### Solution\n\n\n\nOur overall solution is a 3 stage pipeline with additional stuff on the side, illustrated above.\n1. Predict fractures the same way as the public kernel. We use a custom 3D-CNN model for sampling frames in the right vertebra, as well as available fractures bounding boxes.\n2. Re-use the same pipeline in 2.5D feeding 5 frames and adding a LSTM head\n3. Switch to study level feeding predicted probabilities for all the slices to a small RNN or 1D-CNN. We also add the probability of the highest confident fracture box of a yolo-v5-l. \n\n\n#### Comments\n\n- The 2nd model was quite tricky to make work, and we did not have the time to use a better cropping than the 384 center crop, but we had the csv with vertebrae centered crops ready. Joining late is tough.\n- 3D models did not really work for us, but probably because they required a lot of tricks to function optimally.\n- Our models don't even look great at detecting fractures, but the study-level model does a good job exploiting the signal\n- I was optimizing the wrong metric (because of normalization) until the last day, switching to the correct implementation helped quite a lot \n\nThanks for reading, hope to see you all on the next medical imaging comp =)",
"votes": 33
},
{
"id": 2011857,
"postDate": "2022-10-31T21:21:04.063Z",
"content": "A hearty congratulation on 18th place! @theoviel
", "rawMarkdown": "A hearty congratulation on 18th place! @theoviel ", "votes": 1 }, { "id": 2007881, "postDate": "2022-10-28T15:02:43.293Z", "content": "Did you fine tune the yolo-v5 with the bounding box?
", "rawMarkdown": "Did you fine tune the yolo-v5 with the bounding box?", "votes": 1, "replies": [ { "id": 2009845, "postDate": "2022-10-30T11:12:19.213Z", "content": "Yes, exactly
", "rawMarkdown": "Yes, exactly" } ] }, { "id": 2007667, "postDate": "2022-10-28T12:06:36.670Z", "content": "Congrats on 18th place!
\nI tried a similar pipeline but couldn't make it work, can you comment on how much performance improvement you got by adding the yolov5-l probability? @theoviel
Thx !
\nWe got +0.02 CV/LB
A hearty congratulation on 18th place! @theoviel
", "votes": 1, "replies": [] }, { "id": 2007881, "author_name": "Pierre Tisseur", "author_url": "", "post_date": "2022-10-28T15:02:43.293000", "content": "Did you fine tune the yolo-v5 with the bounding box?
", "votes": 1, "replies": [ { "id": 2009845, "author_name": "Theo Viel", "author_url": "", "post_date": "2022-10-30T11:12:19.213000", "content": "Yes, exactly
", "votes": 0, "replies": [] } ] }, { "id": 2007667, "author_name": "Yerram Varun", "author_url": "", "post_date": "2022-10-28T12:06:36.670000", "content": "Congrats on 18th place!
\nI tried a similar pipeline but couldn't make it work, can you comment on how much performance improvement you got by adding the yolov5-l probability? @theoviel
Thx !
\nWe got +0.02 CV/LB
\n\nOur overall solution is a 3 stage pipeline with additional stuff on the side, illustrated above.\n1. Predict fractures the same way as the public kernel. We use a custom 3D-CNN model for sampling frames in the right vertebra, as well as available fractures bounding boxes.\n2. Re-use the same pipeline in 2.5D feeding 5 frames and adding a LSTM head\n3. Switch to study level feeding predicted probabilities for all the slices to a small RNN or 1D-CNN. We also add the probability of the highest confident fracture box of a yolo-v5-l. \n\n\n#### Comments\n\n- The 2nd model was quite tricky to make work, and we did not have the time to use a better cropping than the 384 center crop, but we had the csv with vertebrae centered crops ready. Joining late is tough.\n- 3D models did not really work for us, but probably because they required a lot of tricks to function optimally.\n- Our models don't even look great at detecting fractures, but the study-level model does a good job exploiting the signal\n- I was optimizing the wrong metric (because of normalization) until the last day, switching to the correct implementation helped quite a lot \n\nThanks for reading, hope to see you all on the next medical imaging comp =)",
"2011857": "A hearty congratulation on 18th place! @theoviel ",
"2007881": "Did you fine tune the yolo-v5 with the bounding box?",
"2007667": "Congrats on 18th place! \nI tried a similar pipeline but couldn't make it work, can you comment on how much performance improvement you got by adding the yolov5-l probability? @theoviel "
}
}