Weighted Focal Loss: An Effective Loss Function to Overcome Unbalance Problem of Chest X-ray14

Abstract

X-ray film has been applied widely in the elementary diagnosis and screening of thorax diseases. The rapid development of deep learning has improved automatic diagnosis of thorax disease. But different thorax disease in database inevitably exist the problem of unbalanced quantity because of the different morbidity in real life. So, the restricted factor is not just the architecture of the models but the unbalanced problem of database. What's more, different from the same problem of natural images, this unbalanced problem is not just the unbalanced quantity but also the unbalanced complexity of diagnosis between samples. In this study, we introduced the Focal Loss and proposed the W-FL (weight Focal Loss) function which makes the model pay more attention to the difficult samples and those kinds of diseases which have small quantity and protect the loss and computed gradients from overwhelming by easy negatives. We evaluate the efficiency of W-FL function on the public Chest X-ray14 database using prevailing per-trained models including AlexNet, Inception-v3 and ResNet. We exceed at least 2% average AUC value on 14 different thorax diseases than the W-CEL function which is commonly proposed to solve the unbalanced problem in this database.