UCLA

Magnetic resonance imaging (MRI) is a powerful imaging technique for visualizing soft tissues and characterizing tissue properties. It can be used for the assessment of meniscal, ligamentous, and cartilaginous lesions in the knee. When the extracellular matrix of articular cartilage is compromised, water moves more freely within the cartilage, which leads to prolonged T2 relaxation times. Consequently, quantitative T2 maps have been used to understand the pathophysiology of osteoarthritis or follow-up monitoring of knee cartilage after surgery. Standard T2 quantification techniques usually use spin echo-based sequences which require long acquisition times. Three-dimensional (3D) dual echo steady state (DESS) MRI has shown its potential for time-efficient T2 mapping in the knee. DESS acquires images with distinct contrasts (“FID” and “Echo”). This feature of DESS allows for both morphological T2-weighted (T2w) imaging and quantitative T2 mapping in a single scan. Our research group at UCLA has developed a protocol to acquire 3D DESS scans with isotropic high-resolution. However, the scan time still needs to be reduced to facilitate eventual translation. In order to reduce acquisition time in T2 mapping sequences, previous works developed methods to reconstruct images using undersampled data. Previous proposals such as compressed sensing (CS) reconstruction have been used to reconstruct images from undersampled data while mitigating undersampling artifacts. However, CS reconstruction methods are time-consuming and the results are sometimes overregularized. Deep learning (DL) based image enhancement or reconstruction methods provide a solution to the shortcomings of CS reconstruction by learning the mapping between undersampled input data and high-quality output images from large reference datasets, and providing rapid inference times. Since the isotropic high-resolution 3D DESS acquisition designed by our group at UCLA is relatively recent and has a limited number of datasets, strategies such as transfer learning, i.e., pretraining on a larger available reference dataset, may be needed to obtain a DL network that can produce high-quality output images from our undersampled high-resolution 3D DESS MRI data. Another main challenge of DL-based image enhancement or reconstruction methods is that it is difficult to understand why/how they work or to predict its performance. In medical image enhancement, image fidelity must be prioritized. Obstruction or elimination of image details may confound diagnostic decisions. To overcome this problem, recently there are some works on developing DL networks with uncertainty estimation to predict error in enhancement tasks. Thus, the aim of this work is to shorten the scan time of 3D DESS MRI by reconstructing high quality images from undersampled data using DL networks that incorporate uncertainty estimation and transfer learning.