Hybrid neural network classification of knee osteoporosis using X-ray images and clinical features

Abstract

Osteoporosis is a degenerative disease marked by decreased bone density and increased fracture risk, particularly in the knee. Early detection is crucial, but conventional diagnostic methods are often subjective and require specialized skills. This study developed an artificial intelligence (AI) model for automatic osteoporosis classification using a hybrid method combining knee X-ray images and clinical data, integrating a convolutional neural network (CNN) for image features with an artificial neural network (ANN) for tabular data. The dataset comprised 239 patients with knee X-ray images and 27 clinical parameters. Pre-processing included handling missing values, encoding, and selecting the 20 best features via the analysis of variance (ANOVA) F-test, with T-score as the most influential parameter. Contrast-limited adaptive histogram equalization (CLAHE) improved image quality, while the synthetic minority oversampling technique (SMOTE) and image augmentation produced a balanced dataset of 462 samples. The hybrid model achieved 97.87% accuracy in classifying normal, osteopenia, and osteoporosis cases. Benchmarked against five alternative backbones (residual network (ResNet) 50V2, densely connected convolutional network (DenseNet) 121, mobile network (MobileNet) V2, vision transformer (ViT), and swin transformer) under the same framework, only the swin transformer matched this performance, while the others reached 74.47–82.98% accuracy, confirming the proposed backbone's contribution. These results show that combining imaging and clinical data improves diagnostic accuracy and supports AI-based early detection in orthopedics.