search

eureka7mt / chexnet-mxnet

A mxnet implementation of CheXNet
MIT License
8 stars 6 forks source link

would you mind posting your results? #1

Open yian2271368 opened 6 years ago

yian2271368 commented 6 years ago

hey, I have tested some images using demo.py, the results are always 'no findings in :'. Thus, I am wondering if you can post the results(accuracy for each disease). I will really appreciate you can help. Thanks

eureka7mt commented 6 years ago

Hi @yian2271368 .The average AUROC is about 0.84. In the network,the output function is not a softmax function,so you should choose a proper threshold for best prediction performance.High AUROC only guarantee that you could find a threshold that predict well.BTW, it's a global threshold for all prediction, I will use 14 threshold for each prediction instead of it soon. Finally,I don't suggest that you use the CheXNet to diagnose according to Luke's blog