I am not sure if this problem is unique to me and I have not succeeded at overcoming it yet, but I feel like there is a problem when training with the Dicom files, there are only 176 in the train set, and I think models could just learn the relation without any generalization.
\nHandling the files as slices with 2d CNN's, the CV error suffers significantly, even with models as small as MobileNetv2.
\nUsing a VAE might help with this problem but I would guess not. also using 3d models I feel would increase the overfitting as it would increase the dimensionality of the data, although with an autoencoder you can train on external data its unsupervised.
\nThe best score I have achieved so far is using purely tabular and the CV decreases when I introduce a CNN.
\nI wonder if anyone is running into the same wall and if they use heavy data augmentation, dropout, regularization or smaller models :)
\nBest of luck to everyone.
", "messages": [ { "id": "978052", "postDate": "08/19/2020 22:34:59", "content": "I am not sure if this problem is unique to me and I have not succeeded at overcoming it yet, but I feel like there is a problem when training with the Dicom files, there are only 176 in the train set, and I think models could just learn the relation without any generalization.
\nHandling the files as slices with 2d CNN's, the CV error suffers significantly, even with models as small as MobileNetv2.
\nUsing a VAE might help with this problem but I would guess not. also using 3d models I feel would increase the overfitting as it would increase the dimensionality of the data, although with an autoencoder you can train on external data its unsupervised.
\nThe best score I have achieved so far is using purely tabular and the CV decreases when I introduce a CNN.
\nI wonder if anyone is running into the same wall and if they use heavy data augmentation, dropout, regularization or smaller models :)
\nBest of luck to everyone.
", "rawMarkdown": "I am not sure if this problem is unique to me and I have not succeeded at overcoming it yet, but I feel like there is a problem when training with the Dicom files, there are only 176 in the train set, and I think models could just learn the relation without any generalization. \n\nHandling the files as slices with 2d CNN's, the CV error suffers significantly, even with models as small as MobileNetv2.\n\nUsing a VAE might help with this problem but I would guess not. also using 3d models I feel would increase the overfitting as it would increase the dimensionality of the data, although with an autoencoder you can train on external data its unsupervised.\n\nThe best score I have achieved so far is using purely tabular and the CV decreases when I introduce a CNN.\n\nI wonder if anyone is running into the same wall and if they use heavy data augmentation, dropout, regularization or smaller models :)\n\nBest of luck to everyone.", "votes": null }, { "id": "980891", "postDate": "08/22/2020 01:19:18", "content": "I am assuming that only small features in the images are important for prediction. So I am taking patches from a single CT slice, selecting small features from that and trying to improve score. Will post the notebook if I get any improvement at all.
", "rawMarkdown": "I am assuming that only small features in the images are important for prediction. So I am taking patches from a single CT slice, selecting small features from that and trying to improve score. Will post the notebook if I get any improvement at all.", "votes": null }, { "id": "981401", "postDate": "08/22/2020 12:22:12", "content": "I am having the same trouble, i am trying to make the training easier for 3d CNN, regularizers, data augm, img preprocessing …. few improvements but still not overtook the simple baselines ….
", "rawMarkdown": "I am having the same trouble, i am trying to make the training easier for 3d CNN, regularizers, data augm, img preprocessing .... few improvements but still not overtook the simple baselines ....", "votes": null } ], "comments": [ { "id": 980891, "author_name": "ajenningsfrankston", "author_url": "", "post_date": "08/22/2020 01:19:18", "content": "I am assuming that only small features in the images are important for prediction. So I am taking patches from a single CT slice, selecting small features from that and trying to improve score. Will post the notebook if I get any improvement at all.
", "votes": null, "replies": [] }, { "id": 981401, "author_name": "enric1296", "author_url": "", "post_date": "08/22/2020 12:22:12", "content": "I am having the same trouble, i am trying to make the training easier for 3d CNN, regularizers, data augm, img preprocessing …. few improvements but still not overtook the simple baselines ….
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "978052": "I am not sure if this problem is unique to me and I have not succeeded at overcoming it yet, but I feel like there is a problem when training with the Dicom files, there are only 176 in the train set, and I think models could just learn the relation without any generalization. \n\nHandling the files as slices with 2d CNN's, the CV error suffers significantly, even with models as small as MobileNetv2.\n\nUsing a VAE might help with this problem but I would guess not. also using 3d models I feel would increase the overfitting as it would increase the dimensionality of the data, although with an autoencoder you can train on external data its unsupervised.\n\nThe best score I have achieved so far is using purely tabular and the CV decreases when I introduce a CNN.\n\nI wonder if anyone is running into the same wall and if they use heavy data augmentation, dropout, regularization or smaller models :)\n\nBest of luck to everyone.", "980891": "I am assuming that only small features in the images are important for prediction. So I am taking patches from a single CT slice, selecting small features from that and trying to improve score. Will post the notebook if I get any improvement at all.", "981401": "I am having the same trouble, i am trying to make the training easier for 3d CNN, regularizers, data augm, img preprocessing .... few improvements but still not overtook the simple baselines ...." }, "source": "meta" }