Data-Driven Learning Framework for Fast Quantitative Knee Joint Mapping

PROJECT SUMMARY Osteoarthritis (OA), a leading cause of chronic disability in the elderly population, occurs with the degradation of the extracellular matrix of articular cartilage, mainly composed of proteoglycan, collagen fibers, and water. Early diagnosis of cartilage degeneration requires the detection of changes in proteoglycan concentration and collagen integrity, preferably non-invasively and before any morphological changes occur. Spin-spin relaxation time (T2) and spin-lattice relaxation time in the rotating frame (T1ρ) can provide quantitative information about the structure and biochemical composition of the cartilage before morphological changes occur. Mono-exponential (ME) models can characterize the T2 and T1ρ relaxation processes and map it for articular cartilage in the knee joint. A recent meta-analysis showed that T1ρ provides more discrimination than T2 for OA. However, the ME model alone cannot provide distinct information from different compartments of the cartilage. Recent studies have shown that T1ρ relaxation might have bi-exponential (BE) components, following the hypothesis of the multi- compartmental structure of the cartilage. BE T2 relaxation has shown better diagnostic performance than ME for OA and can show the dispersion of the relaxation times, reflecting the heterogeneity in the macromolecular environment of water in the cartilage. BE analysis of cartilage typically requires a larger number of acquisitions with different spin-lock times (TSLs) or echo times (TEs), resulting in long scan time. High spatial resolution is also needed to visualize the thin and curved cartilage and fine structures in the knee joint. As a result, in vivo application of BE three-dimensional (3D) T1ρ and T2 mapping techniques is still very limited. Compressed sensing (CS) combined with parallel imaging (PI) can accelerate acquisition and reduce the scan time required for ME 3D T1ρ and T2 mappings. T1ρ scans can be reduced from 30 min to ~3 min with an error smaller than 6.5%. However, the error is two to three times larger for BE mapping. This problem can be potentially solved by optimizing the sampling times (TSLs for T1ρ and TEs for T2) and the free parameters of the CS approach (k- space sampling pattern, regularization function, regularization parameter, and minimization algorithm parameters) using fully sampled 3D knee joint datasets, supported by machine learning tools. The overarching goal of this proposal is to develop, optimize, and translate a high-spatial-resolution, rapid 3D magnetic resonance imaging sequence using data-driven learning-based CS for assessment of the human knee joint and using ME and BE 3D T1ρ (T2) mapping for improved biochemical characterization of cartilage and menisci on a standard clinical 3T scanner.