Semi-Supervised Medical Image Segmentation with Dual-View Differential Feature Reinjection.

Abstract

The growing availability of multi-institutional medical imaging data, facilitated by emerging 6G networks, enables large-scale 3D segmentation studies. However, voxel-wise annotation remains costly and scarce, and existing semi-supervised methods, which rely mainly on output-level consistency, often neglect semantic cues in encoder representations, resulting in weak supervision in uncertain regions and over-smoothed boundaries. To address these challenges, we propose a Dual-View Differential Feedback (DVDF) semi-supervised framework for 3D medical image segmentation, which leverages a feature-reinjection-guided feedback loop to transform prediction discrepancies in uncertain regions into learnable supervisory signals. Additionally, we propose a Semantic Refinement Decoder (SRD) to enhance decoder consistency and capture fine details by aligning semantic information across layers and reconstructing features at multiple scales. To further enhance robustness, we incorporate a Semantic Aggregation Attention (SAA) module at the bottleneck, which aggregates long-range contextual information and establishes a cross-scale global semantic prior, thereby enabling more effective training on complex structures and ambiguous boundaries. Experiments on the Left Atrium (LA) and Pancreas-CT datasets demonstrate consistent improvements across key metrics under low-annotation settings, validating the effectiveness and robustness of the DVDF framework.