CSCADA: Cycle and Semantic Consistency Adversarial Domain Adaptation for Cross-Modality Medical Image Segmentation

Abstract

Towards a challenging problem cross-modality unsupervised domain adaptation between unpaired CT and MRI images, we propose a generic framework CSCADA, composed of an evolutionary GAN (termed as Cycle-Y-GAN) and an organ segmentor, optimized by adversarial learning manner. Cycle-Y-GAN aims to learn translation knowledge from one paired base (between CT and second MRI domain) and generalize to target unpaired ones to mitigate domain mismatch in image appearance level. The segmentor targets to segment concerned regions and meanwhile close domain shift further in semantic level. Experiment results demonstrate that our method recovers the degraded generalization performance from 20.06% to 70.69% in terms of Dice Similarity Coefficient (DSC) on our dataset.