Lightweight Multi-Branch Network Classifies Bone Fractures Across Radiograph Datasets

Researchers have developed FMC-Net, a task-adaptive lightweight convolutional neural network engineered for two-dimensional fracture classification in radiographic imaging, demonstrating superior overall performance across multiple independent benchmarks. FMC-Net combines lightweight convolutional processing with multi-stage feature extraction, achieving competitive macro F1-scores across multiple fracture classification datasets while preserving computational efficiency.

In clinical orthopaedic practice, radiographic indicators of skeletal fracture frequently present as subtle, localized disruptions in cortical continuity, faint linear structural interruptions, or localized abnormalities situated near anatomical joints. Reliably capturing these pathological markers requires an image analysis architecture capable of modeling sharp local edges, directional bone textures, and multi-scale semantic patterns simultaneously.

To achieve this dual capability, FMC-Net incorporates a specialized multi-branch convolutional architecture comprising six complementary design elements. The framework employs edge-preserving input representations alongside multi-branch residual convolutional blocks to capture diverse feature granularities. For spatial reduction, the system introduces boundary-preserving downsampling mechanisms that prevent the loss of fine skeletal margins during intermediate network stages. Furthermore, the architecture integrates directional strip convolutional attention and global context calibration modules, enabling the model to emphasize elongated, directional bone structures while calibrating global anatomical features. Multi-source classification heads then synthesize shallow-layer edge responses, mid-layer local structural cues, directional textures, and high-level semantic representations into final diagnostic predictions.

The investigative team evaluated FMC-Net across three distinct public fracture X-ray datasets: FracAtlas, MFXR, and OrthoFrac-XR. To thoroughly validate the network against existing technological approaches, the authors compared its classification efficacy against standard efficient convolutional neural networks, vision Transformer architectures, and hybrid CNN-Transformer models. To verify that performance advantages were reliable rather than artifacts of stochastic initialization, training and evaluation procedures were repeated across multiple random seeds. The evaluation framework examined not only overall classification accuracy but also model stability, spatial interpretability, parameter efficiency, and runtime computational requirements.

In experimental evaluations, FMC-Net achieved superior overall classification performance when assessed separately across each of the three benchmark datasets. Specifically, the network recorded macro F1-scores of 0.7783 on FracAtlas, 0.8141 on MFXR, and 0.8093 on OrthoFrac-XR. While overall performance remained stable, detailed category-level analyses uncovered distinct performance characteristics across the different image collections. Within the FracAtlas benchmark, accurate identification of the fracture class represented the primary performance-limiting factor. Conversely, evaluations on the MFXR dataset showed that recognition of the anomaly class outperformed the good class, whereas evaluations conducted on OrthoFrac-XR demonstrated closely matched diagnostic metrics across both evaluation classes.

To examine the internal decision-making behavior of the network, the researchers conducted spatial interpretability analyses using Gradient-weighted Class Activation Mapping, known as Grad-CAM. Visualizations revealed that the model’s primary activation regions aligned closely with relevant anatomical structures, focusing specifically on bony cortical margins, regions adjacent to skeletal joints, and localized zones exhibiting suspected radiographic anomalies. Complementary ablation experiments validated the architectural necessity of the multi-source classification head, multi-branch residual convolutional blocks, and multi-branch downsampling components, demonstrating that each contributed significantly to overall diagnostic accuracy.

Beyond classification performance, efficiency analyses demonstrated that FMC-Net achieves an effective operational balance among total model parameters, floating-point operations (FLOPs), and inference speed. The researchers concluded that FMC-Net delivers a robust, task-adaptive convolutional solution capable of handling two-dimensional fracture X-ray evaluation across variable imaging conditions. By combining competitive accuracy, verified spatial interpretability across bony structures, and low computational overhead, the lightweight model offers a practical technological foundation for orthopaedic computer-aided diagnostic workflows.

Subject of Research: Lightweight convolutional network for fracture classification in two-dimensional radiographs

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

Article References: Li, C., Chen, P., Liu, Q., & Ling, C. (2026). FMC-Net: an anatomy-aware multi-cue lightweight network for fracture classification in radiographs. Frontiers in Medicine, 13, Article 2034134. https://doi.org/10.3389/fmed.2026.2034134

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.2034134

Keywords: fracture classification, radiographs, x-ray imaging, FMC-Net, lightweight neural network, multi-branch convolution, orthopaedic diagnosis, computer-aided diagnosis, cortical continuity, directional strip convolutional attention, boundary-preserving downsampling, global context calibration, FracAtlas, MFXR, OrthoFrac-XR, Grad-CAM, macro F1-score, ablation study