MST-UDA: A Unsupervised Domain Adaptation Framework via 2.5D Multi-Style Perceptual Translation Network and Self-Filtering for Cross-Modal Multi-Organ Segmentation

Zengmin Zhang, peng yanjun, Xiaoning Zhang, Haoran Li

Keywords: Unsupervised domain adaptation, Cross-Modal Segmentation, Style Translation

TL;DR: We propose MST-UDA, a framework using multi-style translation and self-filtering for cross-modal abdominal segmentation, achieving 80.77% DSC for MRI and 62.83% for PET using only CT annotations.

Abstract: Abdominal medical image segmentation is essential for clinical diagnosis and treatment planning. Although abdominal CT multi-organ segmentation has achieved significant progress, MRI and PET modalities face challenges of annotation scarcity and cross-modal domain gaps. Unsupervised domain adaptation (UDA) provides an effective solution. However, existing methods suffer from domain shift and training instability when adapting to multiple target styles. Thus, we propose a UDA framework for MRI/PET abdominal multi-organ segmentation based on multi-style perceptual translation and self-filtering. Achieved accurate segmentation of unlabeled MRI/PET only using CT annotations. Specifically: (1) We designed an enhanced 2.5D Multi-Style Perceptual Translation Network (MST-Net) synthesizing diverse fake MRI/PET images from CT; (2) We train a dense segmentation model using multi-style data to generate pseudo-labels for real MRI/PET images; (3) We filter fake images and pseudo-labels through accuracy and stability assessment to improve final train data quality; (4) Final, we employ a two-stage lightweight segmentation model for accurate and efficient MRI/PET segmentation. Experiments on FLARE2025 validation set show our method achieves excellent performance with fast, low-resource characteristics: MRI and PET average DSC reach 80.77\% and 62.83\%, with 3.47s average inference time and 2479MB peak memory consumption.

Submission Number: 5