You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi All,
I would like some advice. So I am trying to emulate the results of this paper, and I am training the patch classifier right now. I am extracting 256x256 size patches from 1156x892 sized images (Image resizing was done using PIL). There is patient level separation between test and train data. So, 67% of patients are in the training set, and 33% are in the testing set.
Somehow, the Resnet50 is overfitting severely even after data augmentation. It is somehow not learning, and just fitting on the training data.
Any idea as to why this might be happening?
The text was updated successfully, but these errors were encountered:
@oxinrong Its surprising that despite my patch classifier not performing as well as the paper states, on an image level, I was able to achieve similar AUC as the paper! And to answer your question, no I did not find a solution yet
@Neo96Mav You said that you can almost achieve the similar AUC values to the results in the paper, however, I test the trained model with DDSM and the AUC value is around 0.7, which is much lower than expected. Did you also test on DDSM? Did you use some tricky method to preprocessing the dicom images?
Hi All,
I would like some advice. So I am trying to emulate the results of this paper, and I am training the patch classifier right now. I am extracting 256x256 size patches from 1156x892 sized images (Image resizing was done using PIL). There is patient level separation between test and train data. So, 67% of patients are in the training set, and 33% are in the testing set.
Somehow, the Resnet50 is overfitting severely even after data augmentation. It is somehow not learning, and just fitting on the training data.
Any idea as to why this might be happening?
The text was updated successfully, but these errors were encountered: