Resolution Enhancement and Contrast Harmonization for MR Neuroimaging

Project Summary/Abstract Magnetic resonance (MR) neuroimaging is an essential tool for clinical diagnosis, monitoring, and management and for research on the pathophysiology, progression, and treatment efficacy of many neurological diseases. Because MR hardware and software can be different between sites and can change over time as technology improves, it has been extremely difficult to standardize the appearance of MR neuroimages. As a result, establishing consistent, reliable postprocessing for quantitative analysis of these images has been problematic, and there is evidence that radiologists’ assessments are also affected by this variability. Although many research protocols have been able to use standardized acquisitions, clinical acquisition protocols continue to be highly variable because of differences in hardware and software and from local preferences. To reduce costs, for example, although research protocols strive to acquire images using three-dimensional (3D) protocols, clinical protocols commonly acquire many scans using multi-slice, two- dimensional (2D) protocols which have thick slices and sometimes slice gaps. Also, different imaging centers may acquire different sets of tissue contrasts (e.g., T1-weighted and T2-weighted contrasts) with a different set of 3D versus multi-slice 2D acquisitions and may sometimes omit certain tissue contrasts entirely. Modern medical image analysis algorithms have been primarily designed for research protocols—with standardized, high-quality images—and they prove to be highly inconsistent when applied to clinical-quality images. There is a dire need to find a way to use modern AI algorithms on clinical images. This will do two things: first, it will permit translation of existing algorithms to the clinic; and second, it will open the vast catalog of clinical images to be used in training new algorithms. This research program proposes the Resolution Enhancement And Contrast Harmonization (REACH) algorithm which super-resolves multi-slice 2D MRI and adjusts the contrast of both multi-slice 2D and 3D MRI for use in both downstream processing and clinical diagnosis. Specifically, we will: 1) Develop and extensively train an interpretable deep learning algorithm called HACA3+ for harmonization, restoration, and imputation; 2) Develop SMORE+, a super-resolution method that incorporates high-resolution reference images and is computationally fast; 3) Develop and evaluate REACH for resolution enhancement and contrast harmonization; 4) Carry out a radiologist observer study using REACH. REACH addresses several aspects that have not been previously addressed. Most importantly, it will be specifically trained and tested on and for clinical-quality data, addressing issues such as images with thick slices and slice gaps, missing tissue contrasts, and harmonization that is optimized for specific downstream processing methods. We will further demonstrate the potential for REACH harmonization to be used by radiologists for clinical diagnosis. REACH will be made freely available to researchers upon its publication.