Osteoporosis is a widespread metabolic bone disorder characterized by reduced bone mineral density and increased fracture risk, particularly in older adults and postmenopausal women. Early and accurate diagnosis is essential to prevent severe complications. This study proposes a novel osteoporosis detection framework using knee X-ray data, integrating a mathematical feature extraction technique based on the Whirling Triangle. This method captures structural and spatial patterns through geometric relationships such as angles and triangle areas, enhancing feature representation by reducing noise and dimensionality. The extracted features are fused and processed through a hybrid classification architecture comprising an autoencoder and attention-based three Multi-Layer Perceptrons (MLPs) with varying structures. This ensemble network categorizes images into normal, osteopenic, or osteoporotic classes. Extensive experiments with different dataset splits were conducted to validate the model’s generalizability. Comparative analysis with several frontline technique of Deep Learning and Transfer Learning models demonstrates the accuracy and robustness of the proposed approach.