Deep Learning Methods for Harmonization of Heterogeneous Multiple Sclerosis Imaging Data
Abstract
Magnetic resonance (MR) imaging has shown that people with multiple sclerosis (MS) have more white matter lesions in their brains and a higher rate of brain atrophy than do healthy people. Although lesion and brain volumes can be measured using MR images, their values can be quite different when the image acquisition protocol is changed either by using a different scanner or through a change to a given scanner s hardware or software. Unfortunately, it is common for scanner software and hardware to change due to upgrades or routine maintenance, and it is also common for MS patients to visit different facilities for scanning to be performed. Furthermore, clinical trials for evaluating new medications often are performed across multiple centers, which typically involves the use of different scanner hardware and software. Measurements that are intended to measure biological variations are often overwhelmed by the variations caused by the differences in hardware and software at the different centers, and, therefore, the advantages that would be expected by studying a larger population are often diminished or eliminated. The goal of this work is to develop software tools to harmonize--that is, to make more consistent--MR images acquired from MS patients. We will leverage the next generation of machine learning algorithms to enable consistent qualitative assessment and quantification of imaging data, even when the data are acquired with heterogeneous scan properties. Our project has three aims: (1) We will develop and validate algorithms for robust, consistent quantification of brain and lesion volumes, thereby enabling group comparisons and clinical trials involving large numbers of subjects to be less affected by MR scan variations. (2) We will develop and validate algorithms to harmonize scanning changes that may occur over time, enabling subject-specific tracking of volume and intensity changes in brain structure. We will consider scenarios where specialized data are acquired to train our algorithms, as well as when such data are not available. (3) We will test and refine the developed techniques by analyzing MS imaging data acquired at Johns Hopkins University, the National Institutes of Health, and Walter Reed National Military Medical Center. Our team has developed novel machine learning algorithms for mapping the intensity properties between images, so-called synthesis algorithms, which will be used in concert with volumetric analysis methods to produce more robust measurements. We also capitalize on our own recent developments enhancing the spatial resolution of imaging data without the need for training data. Our preliminary results have shown a remarkable ability to extract improved depictions of three-dimensional structure using these methods. By the end of the project, we will have developed a suite of software tools for robust image harmonization and analysis in MS patients that will be freely distributed to the scientific community. We will demonstrate that the harmonized images will improve the ability of clinicians to track changes in brain structure that occur over time. The ability to more accurately characterize brain changes will allow a better-informed selection of treatment paradigms for patients. In addition, we will demonstrate that our techniques will enable pooled analysis of heterogeneous, multi-center imaging data, increasing the ability to detect subtle patterns that otherwise would have been obscured.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 10, 2021
- Source ID
- W81XWH2010912
Entities
People
- Jerry L. Prince
Organizations
- Henry M. Jackson Foundation for the Advancement of Military Medicine
- United States Army