ResCap-MobileNet: Integrating Multi-modal Feature Extraction and Ensemble Learning for Osteoporotic Fracture Detection

Osteoporotic fracture risk prediction algorithms using clinical risk factors, with or without bone mineral density measurements, have improved the accuracy of treatment targeting, yet current methods often suffer from limitations in image quality, computational cost, and diagnostic precision. To address these challenges, this methodology proposes a novel ensemble learning framework that introduces a novel ensemble learning framework called ResCap-MobileNet, to improve osteoporosis diagnosis using multi-modal medical imaging data taken from X-ray images. The approach begins with preprocessing, where contrast-limited adaptive histogram equalization (CLAHE) enhances image quality, followed by guided filtering for denoising and augmentation techniques like shearing and flipping. Segmentation is performed using a dual-scale residual U-Net (DSR_U-Net) for accurate identification of osteoporotic regions. For feature extraction, graph convolution networks (GCNs) preserve node connectivity in irregular data, while spherical harmonic transforms (SHTs) and topological data analysis (TDA) capture spatial and topological patterns. Feature selection is optimized through the self-adaptive hippopotamus optimization (SAHO) algorithm, which enhances global exploration with predator defense and exploration strategies. The classification stage utilizes ResCap-MobileNet, which integrates ResNeXt-101, MobileNetV3, and a capsule network (CapsNet) to ensure precise identification of osteoporotic fractures.