An equal-weight soft-voting CNN ensemble with explainable AI for three-class knee radiograph classification

Osteoporosis is a common metabolic bone disorder associated with reduced bone strength and increased fracture risk. Although Dual-Energy X-ray Absorptiometry (DEXA) is the clinical standard for bone mineral density assessment, its cost and limited availability can restrict its use in resource-constrained settings. This study presents an explainable deep learning framework for three-class radiographic pattern classification using standard knee X-ray images. Two publicly available datasets were combined, yielding 2,186 radiographs labeled as Normal, Osteopenia, and Osteoporosis. A preprocessing pipeline comprising morphological opening, contrast-limited adaptive histogram equalization, and intensity normalization was applied, followed by geometric augmentation that expanded the dataset to 8,744 image observations. Eight pretrained CNN architectures were evaluated, and an equal-weight soft-voting ensemble comprising DenseNet201, MobileNetV2, and InceptionV3 was constructed by averaging their predicted class probabilities. The ensemble achieved an accuracy of 98.17% on the augmented experimental test partition and achieved the highest observed performance among the evaluated individual and ensemble configurations. LIME and Integrated Gradients were employed to provide qualitative visualizations of image regions associated with selected predictions. A prototype web application was also developed for individual and batch image classification. Because the labels were inherited from public radiographic repositories and were not directly linked to verified DEXA-derived measurements, the framework should be interpreted as a repository-label-based radiographic pattern classification system rather than a method for quantitative bone mineral density estimation. Furthermore, the reported results represent internal evaluation on the constructed augmented dataset and do not establish independent patient-level or clinical generalization.