XAI-Enhanced CNN Framework for Osteoporosis Screening Using Simple X-Ray Radiograph

Abstract

Osteoporosis is the most commonly found skeletal disorder, and it leads to an increased fracture rate, disability and mortality by reducing bone density and causing structural bone degradation. Women in the postmenopausal phase are the most susceptible to this disorder. Although DXA is the common standard for bone mineral density evaluation, its limited availability and high expense pose a significant challenge and make it difficult to adopt in routine screening. In this work, an interpretable deep learning pipeline using a fine-tuned ResNet50 is developed for opportunistic osteoporosis screening. Based on T-score categories obtained using DXA, the model categorizes knee images into three types: Normal, Osteopenia, or Osteoporosis. To ensure the uniformity of the images, each image undergoes standardization and normalization first before being augmented using rotation, flipping, and contrast. On the test data, the model delivered a macro-averaged accuracy of 88.0%, precision of 87.7%, recall of 87.2%, F1-score of 87.3%, specificity of 93.3%, and AUC-ROC of 0.926. By making the predictions of the model clinically interpretable using Grad-CAM, important regions such as distal femur, proximal tibia, subchondral bone, and periarticular trabecular regions have been identified. This indicates that the enhanced CNN by XAI can be potentially used for the purpose of pre-screening osteoporosis.