Unsupervised Domain Adaptation for Cross-modality Abdominal Organ Segmentation via Organ Attention Style Transfer and Dual-stage Pseudo Label Filtering

Huamin Wang, Jianghao Wu, Guotai Wang, Xianhao Zhou, Jinlong He

Keywords: Unsupervised domain adaptation, Abdominal organ segmentation, Dual-stage label filtering

Abstract: Unsupervised domain adaptation (UDA) for abdominal organ segmentation from labeled Computed Tomography (CT) to unlabeled Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET) presents a significant challenge due to large cross-modality gaps. To address this, we propose a comprehensive multi-stage framework that synergistically combines structure-preserving image synthesis, a robust segmentation architecture, and an advanced self-training pipeline. Initially, we leverage an Organ Attention CycleGAN to synthesize anatomically-faithful MRI and PET images from labeled CTs. These synthetic images first train a Coarse-to-Fine segmentation network, which is then refined through a sophisticated self-training scheme. This scheme features a novel dual-stage pseudo-label filtering pipeline that first selects plausible samples based on anatomical consistency and then generates high-precision consensus labels via model ensembling. Evaluated on the FLARE 2025 Task 3 validation set, our complete framework achieves a mean Dice score of 81.21\% on MRI and 81.43\% on PET, demonstrating the efficacy of our approach in bridging the domain gap without requiring any target-domain annotations.

Submission Number: 2