Interpretable Deep Learning for Opportunistic Knee Osteoporosis Screening via Multi-Class X-Ray Analysis and ResNet-18 Feature Attention

Abstract

Osteoporosis remains underdiagnosed globally due to limited access to dual-energy X-ray absorptiometry (DEXA), particularly in resource-constrained settings. This study proposes a fully automated, explainable deep learning framework for classifying knee anteroposterior (AP) radiographs into three categories: Normal, Osteopenia, and Osteoporosis. A dataset of 5,400 expertly annotated knee X-rays was used to train a fine-tuned ResNet-18 backbone, supported by advanced preprocessing techniques including CLAHE, anisotropic diffusion filtering, Retinex normalization, and adaptive gamma correction. Automated region-of-interest (ROI) extraction was applied to focus on clinically relevant areas. For interpretability, Grad-CAM visualizations and feature embeddings using PCA, UMAP, and t-SNE were employed to analyze model focus and feature separability. The model achieved 96.1 % accuracy, a macro F1-score of 0.96, and an AUC of 0.95 on an independent test set, significantly outperforming baseline models such as SVM+HOG, VGG-16, and DenseNet-121. The proposed framework is clinically transparent, computationally efficient, and well-suited for opportunistic osteoporosis screening in primary care, particularly in mobile or rural healthcare contexts.

Ähnliche Arbeiten