search

lishen / end2end-all-conv

Deep Learning to Improve Breast Cancer Detection on Screening Mammography
Other
367 stars 126 forks source link

Patch Classifier (Severe Overfitting) #10

Open jain-avi opened 6 years ago

jain-avi commented 6 years ago

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?

oxinrong commented 5 years ago

@Neo96Mav Hello, have you find the solution yet ?

jain-avi commented 5 years ago

@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

lucky777song commented 5 years ago

@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?