SoftMix: A Novel Data Augmentation for Patched Medical Image Classification

Medical images are difficult to collect, so the number of datasets available for training is small. Training on sparse datasets may easily lead to problems such as overfitting, which will backfire on our experiment. Data augmentation can effectively solve this problem in the universe field. However, in the field of medical images, the model need to consider more about the correlation between global features and internal instances information. To some extend, traditional data augmentation methods are limited due to the particularity of medical images. Therefore, we propose a novel data augmentation method with soft mixing strategy in adaptive patch, termed SoftMix. In training phase, to enhance global features, we apply shuffle process on images in the same batch and split each images into patches with different patch sizes, which changes adaptively according to the training loss. Next, we apply mixing operations with the shuffle group on some patch in the same location to enhance internal instance information. Furthermore, the proportion of the constant patches to the image is also dynamically determined by the training loss. Extensive experimental validation demonstrates that our method achieves challenging results on 3 representative medical image classification tasks.