Hello kagglers,
\nI try some data augmentation methods(affine, elastic transformโฆ) and dropout, but it can not help to improve the validation score(cross validation=5 with auc score floats around 0.5), and I use the Resnet, BCELoss and Sigmoid activation function.
And I am still confused about why the method causes overfitting with training data.
\nHow can I try to improve the score?๐
Thanks.
", "messages": [ { "id": "1478729", "postDate": "08/18/2021 05:51:44", "content": "Hello kagglers,
\nI try some data augmentation methods(affine, elastic transformโฆ) and dropout, but it can not help to improve the validation score(cross validation=5 with auc score floats around 0.5), and I use the Resnet, BCELoss and Sigmoid activation function.
And I am still confused about why the method causes overfitting with training data.
\nHow can I try to improve the score?๐
Thanks.
", "rawMarkdown": "Hello kagglers,\nI try some data augmentation methods(affine, elastic transform...) and dropout, but it can not help to improve the validation score(cross validation=5 with auc score floats around 0.5), and I use the Resnet, BCELoss and Sigmoid activation function. \n\nAnd I am still confused about why the method causes overfitting with training data.\nHow can I try to improve the score?๐\n\nThanks.", "votes": null }, { "id": "1479264", "postDate": "08/18/2021 11:22:57", "content": "Did you tried ensembling models each trained on different mri types ?
", "rawMarkdown": "Did you tried ensembling models each trained on different mri types ?", "votes": null }, { "id": "1479350", "postDate": "08/18/2021 12:31:55", "content": "Thanks for your advice. I haven't try ensembling models. Now, I take the experiment only on FLAIR type as I had seen some works are based on FLAIR type work well and the tumors on FLAIR type are easy to be seen. I will train with the rest type after getting auc score higher than 0.65. However, the validation loss and auc score cannot change stably and even lower than 0.62 after training auc greater than 0.9.๐ค
\n\n", "rawMarkdown": "Thanks for your advice. I haven't try ensembling models. Now, I take the experiment only on FLAIR type as I had seen some works are based on FLAIR type work well and the tumors on FLAIR type are easy to be seen. I will train with the rest type after getting auc score higher than 0.65. However, the validation loss and auc score cannot change stably and even lower than 0.62 after training auc greater than 0.9.๐ค\n\n> Chen X, Zeng M, Tong Y, et al. Automatic prediction of MGMT status in glioblastoma via deep learning-based MR image analysis[J]. BioMed Research International, 2020, 2020.", "votes": null }, { "id": "1479446", "postDate": "08/18/2021 13:22:46", "content": "Chen X, Zeng M, Tong Y, et al. Automatic prediction of MGMT status in glioblastoma via deep learning-based MR image analysis[J]. BioMed Research International, 2020, 2020.
\n
I don't think it will be a good approach however I have one recommendation, try training two cnns of same mri type and ensemble them. Mostly it will be one 3d cnn and one 2d cnn.
\nT2w images will also work well along with FLAIR images
๐Got it, I will try it right now. Thank you very much!๐
", "rawMarkdown": "๐Got it, I will try it right now. Thank you very much!๐", "votes": null }, { "id": "1484080", "postDate": "08/21/2021 04:10:45", "content": "Since the dataset is very small overfitting is expected.
\nI would suggest heavy regularisation with early stopping for finding the best result with the model you have.
I tried a approach with resnet and would suggest using fc layers with dropout in place of average pooling.
", "rawMarkdown": "Since the dataset is very small overfitting is expected. \nI would suggest heavy regularisation with early stopping for finding the best result with the model you have.\n\nI tried a approach with resnet and would suggest using fc layers with dropout in place of average pooling.", "votes": null }, { "id": "1484155", "postDate": "08/21/2021 05:14:30", "content": "If small dataset was the problem then think 2d cnns will work well
", "rawMarkdown": "If small dataset was the problem then think 2d cnns will work well", "votes": null }, { "id": "1484191", "postDate": "08/21/2021 05:40:26", "content": "Tho with 2d cnns the problem in my opinion is finding comparable scan
", "rawMarkdown": "Tho with 2d cnns the problem in my opinion is finding comparable scan", "votes": null } ], "comments": [ { "id": 1479264, "author_name": "swaralipibose", "author_url": "", "post_date": "08/18/2021 11:22:57", "content": "Did you tried ensembling models each trained on different mri types ?
", "votes": null, "replies": [ { "id": 1479350, "author_name": "yangqinzhu", "author_url": "", "post_date": "08/18/2021 12:31:55", "content": "Thanks for your advice. I haven't try ensembling models. Now, I take the experiment only on FLAIR type as I had seen some works are based on FLAIR type work well and the tumors on FLAIR type are easy to be seen. I will train with the rest type after getting auc score higher than 0.65. However, the validation loss and auc score cannot change stably and even lower than 0.62 after training auc greater than 0.9.๐ค
\n\n", "votes": null, "replies": [] }, { "id": 1479446, "author_name": "swaralipibose", "author_url": "", "post_date": "08/18/2021 13:22:46", "content": "Chen X, Zeng M, Tong Y, et al. Automatic prediction of MGMT status in glioblastoma via deep learning-based MR image analysis[J]. BioMed Research International, 2020, 2020.
\n
I don't think it will be a good approach however I have one recommendation, try training two cnns of same mri type and ensemble them. Mostly it will be one 3d cnn and one 2d cnn.
\nT2w images will also work well along with FLAIR images
๐Got it, I will try it right now. Thank you very much!๐
", "votes": null, "replies": [] } ] }, { "id": 1484080, "author_name": "aryamansharma47", "author_url": "", "post_date": "08/21/2021 04:10:45", "content": "Since the dataset is very small overfitting is expected.
\nI would suggest heavy regularisation with early stopping for finding the best result with the model you have.
I tried a approach with resnet and would suggest using fc layers with dropout in place of average pooling.
", "votes": null, "replies": [ { "id": 1484155, "author_name": "swaralipibose", "author_url": "", "post_date": "08/21/2021 05:14:30", "content": "If small dataset was the problem then think 2d cnns will work well
", "votes": null, "replies": [] }, { "id": 1484191, "author_name": "aryamansharma47", "author_url": "", "post_date": "08/21/2021 05:40:26", "content": "Tho with 2d cnns the problem in my opinion is finding comparable scan
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1478729": "Hello kagglers,\nI try some data augmentation methods(affine, elastic transform...) and dropout, but it can not help to improve the validation score(cross validation=5 with auc score floats around 0.5), and I use the Resnet, BCELoss and Sigmoid activation function. \n\nAnd I am still confused about why the method causes overfitting with training data.\nHow can I try to improve the score?๐\n\nThanks.", "1479264": "Did you tried ensembling models each trained on different mri types ?", "1479350": "Thanks for your advice. I haven't try ensembling models. Now, I take the experiment only on FLAIR type as I had seen some works are based on FLAIR type work well and the tumors on FLAIR type are easy to be seen. I will train with the rest type after getting auc score higher than 0.65. However, the validation loss and auc score cannot change stably and even lower than 0.62 after training auc greater than 0.9.๐ค\n\n> Chen X, Zeng M, Tong Y, et al. Automatic prediction of MGMT status in glioblastoma via deep learning-based MR image analysis[J]. BioMed Research International, 2020, 2020.", "1479446": "I don't think it will be a good approach however I have one recommendation, try training two cnns of same mri type and ensemble them. Mostly it will be one 3d cnn and one 2d cnn. \nT2w images will also work well along with FLAIR images", "1479467": "๐Got it, I will try it right now. Thank you very much!๐", "1484080": "Since the dataset is very small overfitting is expected. \nI would suggest heavy regularisation with early stopping for finding the best result with the model you have.\n\nI tried a approach with resnet and would suggest using fc layers with dropout in place of average pooling.", "1484155": "If small dataset was the problem then think 2d cnns will work well", "1484191": "Tho with 2d cnns the problem in my opinion is finding comparable scan" }, "source": "meta" }