Weighted Cross-Entropy for Unbalanced Data with Application on COVID X-ray images
Özgür Özdemir, Elena Battini Sönmez
Abstract
Since December 2019 the world is infected by COVID-19 or Coronavirus disease, which spreads very quickly, out of control. The high number of precautions for laboratory access, which need to be taken to contain the virus, together with the difficulties in running the gold standard test for COVID-19, result in a practical incapability to make early diagnosis. Recent advances in deep learning algorithms allow efficient implementation of computer-aided diagnosis. This paper investigates on the performance of a very well known residual network, ResNet50, and a lightweight Atrous CNN (ACNN) network using a Weighted Cross-entropy (WCE) loss function, to alleviate imbalance on COVID datasets. As a result, ResNet50 model initialized with pre-trained weights fine-tuned by ImageNet dataset and exploiting WCE achieved the state-of-the-art performance on COVIDXRay-5K test set, with a top balanced accuracy of 99.87%.