Hello everyone, I've been thinking about using 2D CNN for this competition since I couldn't find any 'efficient' 3D networks for classification.
\nSo the idea is sth like this:
\nTaking the all .dcm MRs in one folder (let's say 00001/FLAIR folder) appending them, and creating a 2D image from these frames. And training those frames with 2D CNN. Would that be helpful? Or any of you tried such thing like this?
Thanks.
", "messages": [ { "id": "1498924", "postDate": "09/01/2021 10:12:25", "content": "Hello everyone, I've been thinking about using 2D CNN for this competition since I couldn't find any 'efficient' 3D networks for classification.
\nSo the idea is sth like this:
\nTaking the all .dcm MRs in one folder (let's say 00001/FLAIR folder) appending them, and creating a 2D image from these frames. And training those frames with 2D CNN. Would that be helpful? Or any of you tried such thing like this?
Thanks.
", "rawMarkdown": "Hello everyone, I've been thinking about using 2D CNN for this competition since I couldn't find any 'efficient' 3D networks for classification.\nSo the idea is sth like this:\nTaking the all .dcm MRs in one folder (let's say 00001/FLAIR folder) appending them, and creating a 2D image from these frames. And training those frames with 2D CNN. Would that be helpful? Or any of you tried such thing like this?\n\nThanks.", "votes": null }, { "id": "1499016", "postDate": "09/01/2021 11:40:50", "content": "If 'Taking the all .dcm MRs in one folder (let's say 00001/FLAIR folder) appending them' means creating a more-than-3-channels 2d image, and train it with a 2d cnn model, then yes, i'm doing something like these as well. I'm not so sure about the performance since LB score has a huge gap with my CV score.
\nOne thing about using a pretrained model is that you need to add a layer to decrease your channels back to 3, which cause the training a bit more difficult.
Yeah that's what I was talking about. So I used Effnetv2-M Torch. But at the inference, my submission got an exception. That's why I started to think if something is wrong. And I think I'm gonna go with 3D CNNs. Thanks for your response though.
", "rawMarkdown": "Yeah that's what I was talking about. So I used Effnetv2-M Torch. But at the inference, my submission got an exception. That's why I started to think if something is wrong. And I think I'm gonna go with 3D CNNs. Thanks for your response though.", "votes": null }, { "id": "1499475", "postDate": "09/01/2021 16:58:02", "content": "i used keras model - my result is 0.618
", "rawMarkdown": "i used keras model - my result is 0.618", "votes": null }, { "id": "1500310", "postDate": "09/02/2021 09:21:16", "content": "I have very briefly probed what does the data says when probing it with 2d cnn. I agree also that there is huge amount of 2d cnn studies and backbones available, compared to 3d. But, the \"time domain\" is interesting. Could you find help for your approach from that: https://www.kaggle.com/experienceinai/aivo-vii-nopea. It include the problem-solving with clearly separated abstraction levels (see pictures and explanation) and the code also for 2d cnn, but also some kind of discussion about \"going toward 3d\" but without 3d cnn.
", "rawMarkdown": "I have very briefly probed what does the data says when probing it with 2d cnn. I agree also that there is huge amount of 2d cnn studies and backbones available, compared to 3d. But, the \"time domain\" is interesting. Could you find help for your approach from that: https://www.kaggle.com/experienceinai/aivo-vii-nopea. It include the problem-solving with clearly separated abstraction levels (see pictures and explanation) and the code also for 2d cnn, but also some kind of discussion about \"going toward 3d\" but without 3d cnn.", "votes": null } ], "comments": [ { "id": 1499016, "author_name": "gdoong", "author_url": "", "post_date": "09/01/2021 11:40:50", "content": "If 'Taking the all .dcm MRs in one folder (let's say 00001/FLAIR folder) appending them' means creating a more-than-3-channels 2d image, and train it with a 2d cnn model, then yes, i'm doing something like these as well. I'm not so sure about the performance since LB score has a huge gap with my CV score.
\nOne thing about using a pretrained model is that you need to add a layer to decrease your channels back to 3, which cause the training a bit more difficult.
Yeah that's what I was talking about. So I used Effnetv2-M Torch. But at the inference, my submission got an exception. That's why I started to think if something is wrong. And I think I'm gonna go with 3D CNNs. Thanks for your response though.
", "votes": null, "replies": [] } ] }, { "id": 1499475, "author_name": "zaakciiru", "author_url": "", "post_date": "09/01/2021 16:58:02", "content": "i used keras model - my result is 0.618
", "votes": null, "replies": [] }, { "id": 1500310, "author_name": "experienceinai", "author_url": "", "post_date": "09/02/2021 09:21:16", "content": "I have very briefly probed what does the data says when probing it with 2d cnn. I agree also that there is huge amount of 2d cnn studies and backbones available, compared to 3d. But, the \"time domain\" is interesting. Could you find help for your approach from that: https://www.kaggle.com/experienceinai/aivo-vii-nopea. It include the problem-solving with clearly separated abstraction levels (see pictures and explanation) and the code also for 2d cnn, but also some kind of discussion about \"going toward 3d\" but without 3d cnn.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1498924": "Hello everyone, I've been thinking about using 2D CNN for this competition since I couldn't find any 'efficient' 3D networks for classification.\nSo the idea is sth like this:\nTaking the all .dcm MRs in one folder (let's say 00001/FLAIR folder) appending them, and creating a 2D image from these frames. And training those frames with 2D CNN. Would that be helpful? Or any of you tried such thing like this?\n\nThanks.", "1499016": "If 'Taking the all .dcm MRs in one folder (let's say 00001/FLAIR folder) appending them' means creating a more-than-3-channels 2d image, and train it with a 2d cnn model, then yes, i'm doing something like these as well. I'm not so sure about the performance since LB score has a huge gap with my CV score.\nOne thing about using a pretrained model is that you need to add a layer to decrease your channels back to 3, which cause the training a bit more difficult.", "1499041": "Yeah that's what I was talking about. So I used Effnetv2-M Torch. But at the inference, my submission got an exception. That's why I started to think if something is wrong. And I think I'm gonna go with 3D CNNs. Thanks for your response though.", "1499475": "i used keras model - my result is 0.618", "1500310": "I have very briefly probed what does the data says when probing it with 2d cnn. I agree also that there is huge amount of 2d cnn studies and backbones available, compared to 3d. But, the \"time domain\" is interesting. Could you find help for your approach from that: https://www.kaggle.com/experienceinai/aivo-vii-nopea. It include the problem-solving with clearly separated abstraction levels (see pictures and explanation) and the code also for 2d cnn, but also some kind of discussion about \"going toward 3d\" but without 3d cnn." }, "source": "meta" }