Voxel-Level Text-Visual Alignment with Discrepancy-Aware Fusion for Semi-Supervised Multi-Organ Segmentation

Multi-organ segmentation in medical imaging remains challenging due to its reliance on costly manual annotations. While semi-supervised learning (SSL) mitigates this by utilizing unlabeled data, current methods lack effective integration of medical domain knowledge and struggle to balance local details with global context. To address these, we propose a novel SSL framework that integrates structured, template-based medical text priors with a complementary dual-branch architecture. Our method synergizes a CNN-based V-Net and a Transformer-based SwinUNETR, effectively capturing both local anatomical details and long-range contextual dependencies. We further introduce a task-aware attention-guided text-visual fusion (ATVF) module that dynamically aligns medical text embeddings with visual features at the voxel level, enhancing semantic discrimination. A discrepancy-aware fusion mechanism is also designed to focus learning on uncertain regions by leveraging inter-branch prediction inconsistencies. Extensive experiments across three public datasets (SegTHOR, MM-WHS, MyoPS) under varying annotation ratios demonstrate competitive performance, especially under low annotation regimes. With only 10% labeled data, it achieves a mean Dice score of 70.86% on SegTHOR, surpassing the second-best solution by 5.8%, and demonstrating notable gains on challenging structures such as the esophagus and right atrium. This work provides a robust and annotation-efficient solution for multi-organ segmentation using structured text prompts, demonstrating strong potential for further evaluation in clinical settings.