Neural Architecture Search for Adversarial Medical Image Segmentation

Adversarial training has led to breakthroughs in many medical image segmentation tasks. The network architecture design of the adversarial networks needs to leverage human expertise. Despite the fact that discriminator plays an important role in the training process, it is still unclear how to design an optimal discriminator. In this work, we propose a neural architecture search framework for adversarial medical image segmentation. We automate the process of neural architecture design for the discriminator with continuous relaxation and gradient-based optimization. We empirically analyze and evaluate the proposed framework in the task of chest organ segmentation and explore the potential of automated machine learning in medical applications. We further release a benchmark dataset for chest organ segmentation.