Lightweight Anatomy-Aware Multi-Class Segmentation of Knee Structures on X-ray Images Using SegFormer

Accurate anatomical delineation in knee X-ray images is critical for the quantitative assessment of osteoarthritis (OA). However, deploying highly accurate segmentation models in clinical environments is often hindered by their massive computational requirements. This study proposes a lightweight, anatomy-aware multi-class segmentation framework based on SegFormer-B0, designed to balance accuracy and efficiency. The hierarchical Transformer encoder captures long-range contextual dependencies critical for resolving overlapping bone structures, while a lightweight decoder maintains computational efficiency. To improve anatomical consistency, a refinement strategy is introduced to enforce structural continuity and boundary integrity, particularly in low-contrast regions. Trained on a clinically validated subset of the Osteoarthritis Initiative (OAI) dataset, the proposed method achieves a mean Dice of 0.9570 and HD95 of 2.1940 pixels with only 3.7123M parameters, outperforming existing lightweight models. The framework also demonstrates real-time capability and improved anatomical plausibility, supporting its use in resource-constrained clinical environments.