Adaptive Feature Selection for Hybrid EfficientNet-Residual Learning with Margin-Based Classification for Bone...

Automated bone-fracture detection from radiographic images remains a challenging task due to variations in anatomical structures, imaging conditions, and class imbalance. While deep convolutional neural networks have demonstrated strong performance in medical image analysis, conventional end-to-end architectures often suffer from limited generalization. To address these limitations, this study proposes a hybrid framework termed Adaptive Feature Selection for Hybrid EfficientNet-Residual Learning with Margin-Based Classification (AFS-HERMS) for automated bone-fracture detection. The proposed approach integrates three complementary components: (i) EfficientNet-B4 for high-level feature extraction, (ii) a residual multilayer perceptron (MLP) for nonlinear feature refinement, and (iii) an adaptive feature selection (AFS) module that adaptively reweights latent feature representations. During the first stage of the framework, EfficientNet-B4, the Residual MLP, and the AFS module are trained using an auxiliary Softmax classification head and cross-entropy loss. Following representation learning, the auxiliary classification head is removed, the neural representation network is frozen, and standardized AFS-weighted embeddings are used to train an independent RBF-SVM classifier. This decoupled two-stage design enables the Softmax and RBF-SVM classifiers to be evaluated using the same learned feature representation. Experiments were conducted on FracAtlas, the Human Bone Fracture Multi-Modal Image Dataset (HBFMID), the Bone Fracture Multi-Region X-ray Data (BFMRD), and MURA-v1.1. On the independent test partitions, the downstream RBF-SVM achieved accuracies of 89.49%, 96.97%, and 99.72% on FracAtlas, HBFMID, and BFMRD, respectively. On MURA-v1.1, the mean test accuracy across the seven anatomical regions was 80.69%. The corresponding ROC-AUC values were 91.13%, 99.94%, 100.00%, and 85.05%, respectively. The HBFMID results were calculated at the crop level, whereas the remaining datasets were evaluated at the image level. These results demonstrate the applicability of the proposed decoupled two-stage framework across multiple bone-fracture datasets, although performance varied across datasets, anatomical regions, and evaluation metrics. Further external and prospective validation is required before clinical deployment.