Unsupervised Machine Learning for Osteoporosis Diagnosis Using Singh Index Clustering on Hip Radiographs

A team of researchers led by Vijaya Kalavakonda from India has published a study on arXiv proposing an unsupervised machine learning approach to automate osteoporosis diagnosis from plain hip radiographs. The method leverages the Singh Index (SI), a semi-quantitative scoring system that assesses trabecular bone patterns in the proximal femur. Traditionally, manual SI calculation is time-consuming and requires specialist expertise. The authors used a custom convolutional neural network (CNN) architecture to extract features from 838 unlabeled hip X-ray images of Indian adults aged 20–70. Various clustering algorithms then categorized the images into six SI grades. Comparative analysis revealed that only two clusters produced high Silhouette Scores, indicating promising classification for distinguishing osteoporotic from non-osteoporotic cases.

The study also identified key challenges: dataset imbalance, image quality issues, and the lack of patient clinical data. To enhance diagnostic accuracy, the authors recommend augmenting X-ray images with clinical metadata and reference images, along with advanced preprocessing techniques. They also suggest exploring semi-supervised and self-supervised learning methods to overcome labeling difficulties in large datasets. This work demonstrates a cost-effective, accessible alternative to dual-energy X-ray absorptiometry (DXA) for mass screening, potentially helping address the growing global burden of osteoporosis among aging populations.