Open mystvearn opened 1 month ago
Thank you for releasing the data and code for reference. I learned from your code and write my own code to run inference on single image. I used your well-trained model for multi-lable multi-disease classification task. The result doesn't seem to be correct as it fails to predict most of images extracted from other dataset. Could you please advice on what could go wrong with my code?
@mystvearn
Here are the results using preprocess without brightness balance, as it may be sensitive to the dataset's mean brightness (alternatives include '2', '4', and '6'), with a threshold of 0.4 on a Mac M3 Pro.
multi-label: DR AMD glaucoma myo rvo LS hyper others test0 :- AMD - - - - - - test1 :- AMD glaucoma - - - - - test2 :- AMD - - - - - - test3 :- - glaucoma - - - - - test4 :- - glaucoma - - - - - test5 :DR - - - - - - - test6 :- - - - - LS - others test7 :- - - - rvo - - - test8 :- AMD - - - - - - test9 :DR - - - - - hyper others
multi-class: test0:myo test1:glaucoma test2:others test3:others test4:others test5:DR test6:LS test7:rvo test8:others test9:rvo
I believe the model achieved some correct classifications for AMD, LS, and DR. I attribute several mistakes, such as with myopia, to the fact that the model was trained on the EDDFS dataset, which has an imbalanced sample distribution. Besides, the model misclassified some cases of glaucoma, suggesting that glaucoma diagnosis may require more fine-grained annotations.
I have added the test_single_img.py into the repo, and updated the training, testing and dataset codes for an improvement.
Thank you for releasing the data and code for reference. I learned from your code and write my own code to run inference on single image. I used your well-trained model for multi-lable multi-disease classification task. The result doesn't seem to be correct as it fails to predict most of images extracted from other dataset. Could you please advice on what could go wrong with my code?
Below are the images that I tested the model on