CD Loss: A Class-Center Based Distribution Loss for Discriminative Feature Learning in Medical Image Classification

Abstract

Accurate classification of medical images plays an essential role in the early diagnosis of disease. Although deep learning has achieved great success in medical images, it remains challenging due to the critical issue of significant intra-class variation and inter-class similarity caused by the diversity of imaging modalities and clinical pathologies. For this purpose, we propose a Class-center based Distribution loss (CD-loss) to guide the deep convolution neural networks (DCNNs) to extract more discriminative features for better classification accuracy. In detail, our CD-loss aims to force the extracted features from medical image data follow the distribution of that extracted from natural image data. That is because in general, there is less significant intra-class variation and inter-class similarity in natural images as that in medical images due to their different imaging mechanisms. In addition, the available medical images are usually very limited, and state-of-the-art (SOTA) metric learning loss functions easily suffer from over-fitting on such small data. On the contrary, our CD-loss extracts discriminative features not only based on the small medical image data but also the large natural image data to reduce over-fitting. To appreciate the performance of the proposed loss, in this experiment, several SOTA metric learning loss functions are used for comparison. The results demonstrate the effectiveness of our method in terms of classification accuracy and F1-scores.