Synergistic Image and Feature Adaptation: Towards Cross-Modality Domain Adaptation for Medical Image Segmentation

• DOI: 10.1609/aaai.v33i01.3301865
• Publication Date: 2019-09-13T22:09:43Z
• Authors: Cheng Chen; Qi Dou; Hao Chen; Jing Qin; Pheng-Ann Heng
• Keywords: Feature (linguistics); Margin (machine learning); Modality (human–computer interaction)

This paper presents a novel unsupervised domain adaptation framework, called Synergistic Image and Feature Adaptation (SIFA), to effectively tackle the problem of shift. Domain has become an important hot topic in recent studies on deep learning, aiming recover performance degradation when applying neural networks new testing domains. Our proposed SIFA is elegant learning diagram which synergistic fusion adaptations from both image feature perspectives. In particular, we simultaneously transform appearance images across domains enhance domain-invariance extracted features towards segmentation task. The encoder layers are shared by perspectives grasp their mutual benefits during end-to-end procedure. Without using any annotation target domain, our unified model guided adversarial losses, with multiple discriminators employed various aspects. We have extensively validated method challenging application crossmodality medical cardiac structures. Experimental results demonstrate that recovers degraded 17.2% 73.0%, outperforms state-of-the-art methods significant margin.