Differentiable Automatic Data Augmentation by Proximal Update for Medical Image Segmentation
Wenxuan He, Min Liu, Yi Tang, Qinghao Liu, Yaonan Wang
Abstract
Dear editor, This letter presents an automatic data augmentation algorithm for medical image segmentation. To increase the scale and diversity of medical images, we propose a differentiable automatic data augmentation algorithm based on proximal update by finding an optimal augmentation policy. Specifically, on the one hand, a dedicated search space is designed for the medical image segmentation task. On the other hand, we introduce a proximal differentiable gradient descent strategy to update the data augmentation policy, which would increase the searching efficiency. Results of the experiments indicate that the proposed algorithm significantly outperforms state-of-the-art methods, and search speed is 10 times faster than state-of-the-art methods.