MFF-Net: Multi-Feature Fusion Network for Knee X-Ray Image Segmentation

Accurate segmentation of the distal femur and proximal tibia from knee radiographs is essential for quantitative assessment of bone morphology and osteoarthritis-related structural changes. However, existing methods remain challenged by weak cortical boundaries, anatomical overlap, information loss during down-sampling, and the limited availability of pixel-level annotations. To address these problems, this paper proposes MFF-Net, a self-supervised multi-feature fusion network (MFF) that integrates hierarchical semantic information with multi-scale boundary cues. First, a ResNet-50 encoder is pretrained on unlabeled training radiographs using contrastive learning to obtain anatomy-relevant representations. During supervised fine-tuning, fixed Gaussian-derivative filters at multiple scales extract complementary cortical boundary responses. A gated semantic-edge fusion decoder then suppresses irrelevant edge information and progressively reconstructs the distal femur and proximal tibia. Experiments were conducted on the Osteoarthritis Initiative (OAI) dataset. MFF-Net achieved a mean Dice coefficient of 93.2% on the held-out test set, outperforming the SOTA methods evaluated in this study by 0.8 percentage points. These results demonstrate the potential of MFF-Net for accurate knee-bone delineation and subsequent quantitative radiographic assessment of knee osteoarthritis.