Attention-weighted hybrid deep feature fusion for detecting the rare bone disease hereditary multiple exostoses in hip radiographs

Hereditary Multiple Exostoses (HME) is a rare skeletal disorder characterized by the growth of multiple benign bone tumors. They often lead to orthopedic complications and malignant transformations. Despite advancements in medical imaging, no automated deep learning framework has been specifically developed for the detection of HME in hip radiographs due to subtle bone surface variations and the limited availability of labeled medical data. In this study, we propose a novel hybrid framework that integrates a lightweight custom convolutional neural network (CNN) with a pretrained MobileNetV2 backbone to exploit both domain-specific and general visual features of the images. To overcome the limited data, we used a unique dataset (70% previously unpublished) enhanced with domain-specific augmentation, which was applied exclusively during the training phase. A key component of our approach is an attention-weighted fusion layer that dynamically balances the contributions of each feature stream, thereby enabling adaptive representation learning. The fused embeddings were evaluated in two complementary ways: classification using a deep head and feature extraction followed by training classical classifiers, including Random Forest (RF), XGBoost (XGB), and Support Vector Machine (SVM). Internal validation using tenfold patient-level cross-validation on the original dataset demonstrated high diagnostic consistency, with mean Area Under the Receiver Operating Characteristic Curve (AUC) values reaching 0.9780. The model was also evaluated using an independent, unseen pediatric dataset. The weighted fusion strategy improved the diagnostic accuracy compared with custom CNN-only and MobileNetV2-only by 8.83%. Statistical tests (paired t-test and Wilcoxon) confirmed the reliability of these gains. Furthermore, Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations highlighted clinically meaningful regions around bone structures, providing interpretability for decision support. We also validated the computational efficiency via cross-platform experiments, in which the training and inference were accelerated on an NVIDIA GeForce RTX 3050 GPU and an edge-capable Jetson Orin NX. The proposed approach offers a promising direction for explainable, resource-efficient, and accurate exostoses detection systems in clinical practice.