Automated quantitative analysis of peri-articular bone microarchitecture in HR-pQCT knee images

There is growing interest in applying HR-pQCT to image the knee, particularly in the study of osteoarthritis, which necessitates the development and validation of novel image analysis workflows. In this work, we present and validate the first fully automated workflow for in vivo quantitative assessment of peri-articular bone density and microarchitecture in the human knee. Bone segmentation models were trained by transfer learning with a large dataset of radius and tibia images (N=2,598) and fine-tuned on a knee image dataset (N=131), atlas-based registration was used to identify medial and lateral contact surfaces, and morphological operations combined these intermediate outputs to generate peri-articular regions of interest (ROIs) for morphological analysis. Accuracy was assessed with an external validation dataset (N=131), where predicted and reference morphological parameters showed excellent correspondence (0.86≤R 2 ≤0.99), with moderate bias present in predictions of subchondral bone plate density (-80 mg HA/cm 3 ) and thickness (+0.15 mm). Precision was assessed with a triple-repeat measures dataset (N=29), where the short-term precision RMS%CV estimates ranged from 0.7% to 3.5% when rigid registration was used to synchronize ROI generation across images.