Development, testing and evaluation of automatic methods for knee cartilage segmentation from Magnetic Resonance Images: application to a cohort of patients affected by medial knee osteoarthritis.

Osteoarthritis (OA) is a debilitating degenerative joint disease that has affected 528 million people globally since 1990. The knee is among the most affected joints, with common symptoms including pain, joint stiffness, and swelling. Currently, there is no definitive cure for this disease; however, its progression can be predicted through morphological investigations of the joint cartilage. The segmentation of the femoral cartilage is fundamental in the process of quantitative (thickness and volume) and qualitative (morphology and relaxometry) analysis. Although manual segmentation is the gold standard, it is very time-consuming, difficult to reproduce, and subject to significant variability. The use of automatic segmentation methods, represents a powerful tool to drastically reduce segmentation times and ensure the reproducibility of the results without being affected by inter/intra-operator variability. This thesis, conducted at the Computational Bioengineering Laboratory of the Rizzoli Orthopaedic Institute in Bologna, aimed to develop, test, and evaluate automatic methods for femoral cartilage segmentation from MRI in subjects with medial knee OA following varus deformity. Automatic segmentation approaches have been developed starting from the open source software pyKNEEr, tested on varus knee subjects free from cartilage lesions and compared. Special attention was posed on method generality, robustness and applicability to real clinical scenarios. An innovative approach based on neural networks (2D U-Net) was also developed and tested on the same cohort and also on subjects with lesions of the articular surfaces. A new workflow for adapting pyKNEEr to the Sagittal Proton Density Cube MRI sequence was developed and validated against manually segmented images by an expert radiologist. The foundation of this workflow was the creation of an Average Atlas to be used as a new reference for image segmentation. A workflow was also dedicated to the calibration and training of the 2D UNet used in both patient cohorts. The quality of the obtained segmentations was evaluated using similarity indices concerning the voxel overlap between masks and ground truth (DICE, Jaccard, Volume Similarity) and by the distances between the segmented surfaces (Hausdorff Distance). The developed workflow increased robustness, generality and led to an improvement in accuracy with respect to the original pyKNEEr software and compared well with the best accuracies published in the relevant literature. The results offered by the UNet neural network further improved the quality of the segmentations and achieved promising results also in the case of patients with cartilage lesions. This result encourages future work on the development of the workflow as it could lead to a possible application of automatic segmentation methods in pathological subjects.