Objective: Histochemical characterization of proteoglycan content in articular cartilage is important for the understanding of osteoarthritis pathogenesis. However, cartilage cells may interfere with the measurement of matrix proteoglycan content in small animal models (e.g. mice and rats) due to the high cell volume fraction (38%) in mice compared to human tissue (~1%). We investigated whether excluding the cells from image analysis affects the histochemically measured proteoglycan content of rat knee joint cartilage and assessed the effectiveness of a deep learning algorithm-based tool named U-Net in cell segmentation.
Design: Histological sections were stained with Safranin-O, after which optical densities were measured using digital densitometry to estimate proteoglycan content. U-Net was trained with 600 annotated Safranin-O cartilage images for exclusion of cells from the cartilage extracellular matrix. Optical densities of the ECM with and without cells were compared as a function of normalized tissue depth.
Results: U-Net cell segmentation was accurate, with the measured cell area fraction following largely that of ground-truth images (average difference: 4.3%). Cell area fraction varied as a function of tissue depth and took up 8-21% of the tissue area. The exclusion of cells from the analysis led to an increase in the analyzed depth-dependent optical density of cartilage by approximately 0.6-1.8% (p < 0.01).
Conclusions: Although the effect of cells on the analyzed proteoglycan content is small, it should be considered for improved sensitivity, especially at the onset of the disease during which cells may proliferate in small animals.
Keywords: Convolutional neural network; U-Net; articular cartilage; chondrocytes; densitometry; optical density; osteoarthritis.