Predicting Osteoporosis Risk from Knee Radiographs and Clinical Features through Deep Learning: A Multimodal Approach

Osteoporosis is a prevalent skeletal disease with high morbidity rates in developing countries due to limited access to gold-standard dual-energy X-ray absorptiometry (DEXA) scanning. This study explores a cost-effective alternative using deep learning to analyze knee radiographs alongside clinical parameters.

We developed a dual-stream approach using 239 knee X-ray images with corresponding clinical data from a public domain dataset collected from a bone mineral density (BMD) assessment camp. Seven convolutional neural network (CNN) architectures were evaluated for image classification. Probability scores from the best-performing CNN were integrated with clinical parameters to create a multimodal dataset. Twelve machine learning algorithms were then tested to develop an optimal classification model for predicting normal BMD, osteopenia, and osteoporosis.

Among the CNN architectures, AlexNet demonstrated superior performance in osteoporosis detection with a recall of 0.83, while Inception V3 achieved the highest overall accuracy (0.60). When integrating CNN-derived probabilities with clinical features, the AdaBoost classifier achieved 72% test accuracy with an area under the curve (AUC) of 0.88. The model showed high sensitivity for osteoporosis detection (0.84) and effective discrimination across diagnostic categories, particularly for normal (AUC 0.96) and osteoporotic cases (AUC 0.87).

The multimodal approach combining deep learning-derived radiographic features with clinical parameters demonstrates promising results for osteoporosis screening. This methodology offers potential for implementation in resource-limited settings where DEXA scans are unavailable.