APGC Net: Unsupervised Cross-Modality Adaptation for Multi-organ Segmentation in TAO via adaptive pseudo-label-guided contrastive learning

Motivation: Clinical assessment of thyroid-associated orbitopathy (TAO) demands precise orbital structure segmentation. The central challenge is that neither pre-contrast (T1) nor post-contrast (T1c) MRI alone provides comprehensive segmentation that covers all TAO-affected organs. Goal(s): The primary objective is to develop an automated segmentation framework that can segment multi-modal MRIs even when complete manual labels are not available for both modalities. Approach: An unsupervised domain adaptation approach is proposed to tackle the challenge via adaptative pseudo-label guided cross-modality contrastive learning. Results: This work achieved significant performance improvements in terms of multiple evaluation metrics. Impact: The impact lies in the technical innovation to ensure consistently accurate segmentation of each organ involved in TAO on multi-modal MRI, which is beneficial to alleviating the burden of manual labeling and reducing observer variability.