Development of image-based method for the quantitative assessment of cartilage damage from CT as a biomarker in osteoarthritis

Preclinical studies, like those for drug development, require quantitative methods to assess osteoarthritis (OA). In OA, cartilage damage and subchondral bone deformations are critical indicators. Therefore, this thesis introduces and validates two innovative frameworks for automated OA assessment, centering on the image analysis of cartilages and subchondral bones. Cartilage is typically invisible in Computed Tomography (CT) scans without an injected contrast agent. The first framework introduced is for the volumetric analysis of cartilage to quantify OA, leveraging CT scans enhanced with a radio-contrast agent developed in our lab that targets type II collagen. We hypothesize that the area difference between CT scans, with and without contrast enhancement, may signify cartilage shape. However, this necessitates precise alignment of such scans. In the initial stage of this work package, an innovative, AI framework without manual annotation is introduced, featuring a unique neural network structure, D-net, designed for precise, automated alignment of CT volumes of mouse tibiae. D-net was developed to address extensive translations and a broad range of rotations without a standard dataset reference template, achieving sub-pixel accuracy in translation and a precision of 1.5 degrees in rotation, outperforming previous methods. The next phase included segmenting the cartilage from aligned CT volumes, utilizing a deep learning segmentation method, the Gate Attention-U-net, for enhanced precision and versatility in cartilage segmentation, achieving an 88% DSC and 9um ASD in the tibia's weight-bearing area. After extracting cartilage shape, we compared the thickness and remaining volume of tibial cartilage in the weight-bearing area with a prior semi-manual method, observing a Pearson correlation coefficient (R value) exceeding 0.8 with the semi-manual method (p<0.0001) and a correlation with OA scores from histological images (p<0.05). This emphasizes our automated method's potential as a valuable OA assessment tool, suggesting our image-based framework's efficacy as an OA assessment tool. Moreover, analysis based on the decay property of voxel-wise contrasted intensity after washout uncovers corresponding changes between half-life time and the early progression of OA. This suggests its ability to bridge the divide between radiological imaging and biochemical markers, hinting at its prospective role as an early OA biomarker. Additionally, an osteophyte deformation tracking framework is also introduced to estimate dynamic morphological changes in subchondral bone across various OA stages, utilizing the newly proposed deep learning-based Residual Aligner Network combined with pre-alignment based on D-net. With a voxel-level average error of 22um in experimental results achieved on in-vivo murine knee CT, this framework substantiates its reliability for estimating and monitoring deformations caused by osteophytes, offering comprehensive insights into OA analysis.