Enhanced osteoporosis screening via multi-output deep learning: Segmentation and classification of metacarpal...

Osteoporosis screening from radiographic images has traditionally relied on isolated methods that fail to capture the complex interplay between structural segmentation and diagnostic classification. This paper introduce OMO-Net, a novel multi-output deep learning architecture that simultaneously performs segmentation and classification on metacarpal radiographs to accurately detect osteoporosis. Unlike conventional approaches, OMO-Net integrates a ResNet-50-based feature extractor with dedicated segmentation and classification branches, enabling the network to localize diagnostically significant regions while performing robust classification. This dual-task framework not only enhances the network's sensitivity to subtle bone density variations but also improves overall diagnostic accuracy, achieving an AUC of 99% and an F1-score of 96.88%. Furthermore, visual interpretability techniques such as Grad-CAM and t-SNE analyses corroborate OMO-Net's capacity to focus on critical metacarpal structures. Our results set a new benchmark in osteoporosis screening, highlighting the potential for integrated deep-learning approaches to transform clinical workflows and diagnostic precision.