A Machine Learning Software Tool for Automatic Segmentation and Analysis of T1p MR Images of Knee Cartilage for OA Diagnosis and Treatment Monitoring

Abstract Osteoarthritis (OA) is a debilitating disease that results in cartilage loss and pain, generating a significant impact on morbidity and healthcare costs worldwide. No drug is currently available for OA. Many therapeutic targets are being pursued by pharmaceutical companies to develop successful disease modifying therapy for OA. However, the major impediment in the development of effective OA therapeutics is the absence of an imaging biomarker, sensitive to soft tissue changes, that can be reliably used as a quantitative measure of OA disease progression, as well as of responses to therapy, for measuring disease modifying effects of drug candidates in clinical trials. 3D T1r MRI has been shown to be sensitive in detecting early cartilage abnormalities in OA patients including the finding that there is an elevation in T1ρ values in early, moderate and advanced stages of OA, when compared with corresponding asymptomatic subjects. Therefore, T1r biomarker has a tremendous potential to be used a primary quantitative endpoint for efficacy in OA clinical trials. However, there are two critical issues for implementing the use of T1r biomarker in OA clinical trials that necessitates measurements of its variability across MRI scanners and development of a software tool for automated segmentation of cartilage images and generation of the corresponding T1r values. We have already addressed the first issue by establishing T1ρ values of healthy subjects across MRI scanners with a maximum reproducibility error of less than 10% at a 95% confidence level, a major first step for an imaging method to be used in clinical trials. We will address the 2nd issue in this proposal where we will develop a deep network and multi-atlas label fusion framework for automatic segmentation of knee cartilage from previously collected T1ρ MR images of 25 healthy volunteers (N=300 data sets). We will then calculate the cartilage mean T1ρ values from all the automatically segmented images and compare these with T1ρ values obtained by manual segmentation in each data set. At the end of this study, we will develop a software tool combining the automated cartilage segmentation method and our algorithm to generate the T1r values for automatic segmentation and analysis of T1r MR images of knee cartilage - enabling patient stratification and monitoring the response to treatment in OA clinical trials to develop disease modifying drugs.