Automatic segmentation of bone-cartilage interface from high-resolution phase-contrast micro-CT images using deep learning

Sammanfattning

The bone–cartilage interface, or calcified cartilage, is a thin but critical transition zone that helps distribute mechanical loads across the synovial joint. However, this interface remains difficult to visualize and segment using conventional techniques due to its structural and textural similarity to the adjacent subchondral bone. Accurate segmentation would enable quantitative analysis of the morphological changes of this interface associated with osteoarthritis. This thesis aims to develop a deep learning pipeline for automating segmentation, specifically using the self-configuring No-New-U-Net (nnU-Net) framework. Models were trained on high-resolution synchrotron phase-contrast micro-CT images of human knee samples, initially using sparse manual annotations from two representative cases. The training set was subsequently expanded through confidence- and uncertainty-filtered pseudo-labeling to improve robustness across a wider range of osteoarthritis histopathological severity grades. Analysis of calcified cartilage thickness in the predicted segmentations was performed and compared between the two groups (i.e., healthy and osteoarthritic). The model achieved high segmentation accuracy [Dice score: 0.97] on a held-out test set consisting of samples spanning three osteoarthritis severity grades, with external validation further demonstrated on human thumb samples [Dice score: 0.97]. The pseudo-labeling strategy yielded consistent improvements in accuracy and reduced variability across the test set. Thickness analysis demonstrated increased mean calcified cartilage thickness in the osteoarthritic group compared with the healthy group. While perfect border delineation remains challenging, the results demonstrate the feasibility of using a deep learning-based pipeline for robust, automated segmentation and morphological analysis of calcified cartilage.