Unsupervised 3D classification and characterisation of osteocyte and chondrocyte lacunae in the murine osteochondral unit

Abstract

In the murine osteochondral unit, morphology of chondrocyte lacunae (CCL) within calcified articular cartilage and osteocyte lacunae (OCL) within the underlying subchondral bone reflects local biomechanical conditions that are altered in disease, making reliable discrimination of both lacunar populations important for studying osteochondral physiology and pathology. Histology is generally considered the gold standard for identifying CCL and OCL. However, it is inherently limited to two-dimensional assessment and a relatively small tissue section, restricting comprehensive characterisation of three-dimensional lacunar structure. Synchrotron-radiation micro-computed tomography (SR-microCT) enables volumetric quantification of lacunae; where classification is typically based on the position of each lacuna relative to the cement line separating calcified cartilage from bone, requiring manual delineation of this convoluted interface. Such approaches are labour-intensive, operator-dependent, and rely on access to specialised synchrotron infrastructure that is not widely available. To address these limitations, we present a fully automated, annotation-free pipeline for classifying OCL and CCL in the mouse osteochondral unit based on desktop micro-computed tomography (microCT), independent of cement line identification. The proposed unsupervised workflow requires neither tissue separation nor labelled training data, enabling objective and scalable discrimination of lacunar populations directly from volumetric image data. Seven morphometric and positional features per lacuna were extracted from SR-microCT and desktop microCT datasets and analysed using Uniform Manifold Approximation and Projection (UMAP). Validation of the pipeline outcome on SR-microCT against manual classification across specimens yielded very-high agreement, with Gwet's AC1, the Matthews correlation coefficient, and balanced accuracy each approaching 0.99, in both healthy and post-traumatic osteoarthritis specimens. Such a workflow can be readily adapted for the investigation of osteochondral interface in various other species.