FMC-Net: an anatomy-aware multi-cue lightweight network for fracture classification in radiographs

Objectives Fracture x-ray classification is an important task in orthopaedic image-aided diagnosis. Fracture-related features often present as local alterations in cortical continuity, fine linear structures, and abnormalities near joints, requiring effective modelling of local edges, directional bone structures, and multi-stage features. This study aims to develop a lightweight network for two-dimensional fracture x-ray classification and evaluate its performance, stability, interpretability, and computational efficiency. Methods We proposed FMC-Net, a multi-branch convolutional network for fracture classification. The model employs edge-preserving inputs, multi-branch residual convolutional blocks, boundary-preserving downsampling, directional strip convolutional attention, global context calibration, and multi-source classification heads to jointly utilise shallow-layer edge responses, mid-layer local structures, directional textures, and high-level semantic information. Experiments were conducted on three fracture x-ray datasets, FracAtlas, MFXR and OrthoFrac-XR, and the model was compared with efficient CNNs, Transformer architectures and hybrid CNN-Transformer models. Training and testing were repeated using random seeds, and model performance was further evaluated using Grad-CAM, ablation studies and efficiency analysis. Results FMC-Net achieved superior overall classification performance in separate evaluations on each of the three public datasets. On FracAtlas, MFXR and OrthoFrac-XR, the macro F 1-scores were 0.7783, 0.8141 and 0.8093, respectively. The model results were generally stable, but category-level performance varied across datasets: fracture class recognition remained the primary limiting factor in FracAtlas, the anomaly class performed better than the good class in MFXR, and the metrics for the two classes were relatively close in OrthoFrac-XR. Grad-CAM visualisations show that the response regions of FMC-Net are primarily distributed across bony structures, areas adjacent to joints, and local regions associated with suspected anomalies. Ablation experiments indicate that the multi-source classification head, multi-branch residual convolutional blocks, and multi-branch downsampling have a significant impact on model performance. Efficiency analysis reveals that FMC-Net maintains a reasonable balance between the params, FLOPs, and inference speed. Conclusions FMC-Net provides a task-adaptive, lightweight convolutional network solution for the classification of two-dimensional fracture x-ray images. The model demonstrates stable and competitive performance across various fracture classification datasets, whilst offering corresponding advantages in terms of spatial response and computational efficiency.