Osteosegnet: A Cross-Domain Few-Shot Learning Model for Knee Osteoarthritis MRI Segmentation

Knee osteoarthritis (OA) is one of the leading causes of mobility impairment in middle-aged and elderly populations. Accurate segmentation of knee MRI is essential for structural analysis, lesion assessment, and disease progression monitoring. However, existing deep learning-based segmentation methods typically rely on large amounts of high-quality annotations, whereas clinical MRI data often suffer from label scarcity and substantial cross-scanner domain shifts, thereby limiting model generalization. To address these challenges, this study proposes OsteoSegNet, a cross-domain few-shot segmentation framework tailored for knee MRI. The model adopts ResNet-50 as a shared encoder and introduces a Feature Transformation (FT) module to mitigate cross-domain feature inconsistency. A Self-Attention (SA) mechanism is incorporated to capture long-range spatial dependencies. Furthermore, a multi-branch(MB) decoder is designed to independently reconstruct four anatomical structures-Femur, Patella, Tibia, and Cartilage-while a cross-branch semantic guidance matrix enforces inter-region structural coherence. Experimental results demonstrate that OsteoSegNet consistently outperforms both fully supervised baselines and state-of-the-art cross-domain few-shot segmentation models in terms of Dice, IoU, and HD95, with the most pronounced improvements observed under the 5-shot setting. These findings indicate that OsteoSegNet effectively alleviates the challenges of annotation scarcity and domain shift, offering a promising solution for knee OA segmentation in real-world clinical environments.