A Novel Approach to Unveil Hidden Stratification in Osteoporosis SR-MicroCT Image Classification

Osteoporosis stems from osteopenia, which silently compromises bone strength, becoming notable just when bone fractures occur. In this context, Deep Learning (DL) has brought advantages through Synchrotron-Radiation micro-Computed Tomography (SR-microCT) image analysis investigating bone health starting from the microscale where the disease originates. However, the macro- to microscale clinical knowledge mismatch and the subtle differences among the osteoporotic states remain an issue. Therefore, we propose a DL method to alleviate the hidden stratification issues featuring a Convolutional AutoEncoder K-mean-based clustering approach to derive macro- to microscale information matching. It then translates the knowledge gained to automatically derive majority voting thresholds that, representative of the clinical sign expressed at the macroscale, allow a ResNet18-based model distinguishing each SR-microCT image. Our approach reaches up to 95.55%, 65.55%, and 73.33% accuracy during healthy, osteopenic, and osteoporotic SR-microCT classifications, while it boosts the per-class average accuracy by up to 28.19% from its standard thresholded counterpart.