Enhanced Osteoporosis Detection in Lumbar Vertebrae Using Hybrid Deep Learning Models on DEXA Images

Osteoporosis is a major global health concern that increases morbidity, mortality, and healthcare expenses. It primarily affects older adults and postmenopausal women. Reduced bone density and an increased risk of fractures are the hallmarks of this common illness, which makes early discovery essential for successful treatment. Dual-Energy X-ray Absorptiometry (DEXA) is commonly utilized for calculating bone density; however, traditional methods often lack precision and efficiency, hindering accurate detection of early signs such as osteopenia. To address these challenges, a novel hybrid deep learning-based approach is introduced for detecting and classifying early signs of osteoporosis in individual lumbar vertebrae from spine DEXA images. The proposed You Only Look Once version 9-based Lumbar Vertebrae Detection (YOLOv9-LVD) model accurately localizes and detects lumbar vertebrae regions, including L1, L2, L3, and L4. The detected lumbar vertebrae are then classified using the Hybrid Model for Lumbar Vertebrae Classification (HMLVC), which integrates Deep Convolutional Neural Network (DCNNet) and Deep DenseNet (DDNet) architectures for precise classification into normal or abnormal categories. Statistical analysis of the classified abnormal lumbar regions, including metrics such as mean, coefficient of variation, entropy, and skewness, facilitates the identification of osteopenia and osteoporosis. The methodology was validated using real-time DEXA images obtained from Chennai Hospital. The YOLOv9-LVD and HMLVC models demonstrated superior performance in detecting and classifying lumbar vertebrae compared to conventional methods, achieving high accuracy and improved sensitivity in identifying early-stage osteoporosis. This advanced methodology offers a promising solution for early and accurate osteoporosis detection from DEXA images, with potential applications in enhancing clinical diagnosis and management strategies.