Not working for Real life xrays

[bijoygit]

I have couple of frontal chest x rays which I know have disease. But when I predict it with the model, it shows very less probability for all the 14 diseases.

It doesn't seem to work on real life X-rays

Please help

[zoogzog]

The goal of this project was only to implement the method described in the original paper. There are no guarantees that the proposed approach would provide good results for each individual image.

It has to be acknowledged that although the proposed approach of pathology classification with DenseNet showed good results, some researchers expressed opinions that the ChestXray14 dataset is not an ideal candidate for conducting evaluation tests.

[QingguoC]

Hello, I know you evaluate the model by the AUROC score, but how do you predict one sample from the trained model? After the last sigmoid activation, do you use 0.5 as the threshold for the classification of each disease, or need to calculate the optimal threshold from the training set? Thanks.