Quantifying cortical lesions in multiple sclerosis MRI datasets using multi-contrast post-processing and deep learning.
Authors
Affiliations (5)
Affiliations (5)
- Buffalo Neuroimaging Analysis Center, Department of Neurology, Jacobs School of Medicine and Biomedical Sciences, University at Buffalo, State University of New York, Buffalo, NY, USA. [email protected].
- Buffalo Neuroimaging Analysis Center, Department of Neurology, Jacobs School of Medicine and Biomedical Sciences, University at Buffalo, State University of New York, Buffalo, NY, USA.
- MS Center Amsterdam, Anatomy and Neurosciences, Vrije Universiteit Amsterdam, Amsterdam Neuroscience, Amsterdam UMC location VUmc, Amsterdam, The Netherlands.
- Genentech Inc., South San Francisco, CA, USA.
- Center for Biomedical Imaging at the Clinical Translational Science Institute, University at Buffalo, State University of New York, Buffalo, NY, USA.
Abstract
Multiple sclerosis (MS) is a chronic neurological disease affecting both white and gray matter of the central nervous system. Despite the well-established involvement of cortical lesions in MS, feasibility limitations in their visualization on typical magnetic resonance imaging (MRI) protocols prevent their evaluation in nearly all clinical trials. Recently, several post-processing methods, including synthetic contrasts and artificial intelligence (AI)-based approaches, have shown potential for enhancing cortical lesion detection on conventional MRI data. These methods have the potential to reanalyze existing clinical-trial data to answer key mechanistic questions about both MS development and about treatment effects. We sought to evaluate the feasibility of combining and extending existing methods into a unified framework for analysis using the data from the large, multicenter, phase 3 ORATORIO trial (full n = 732, age=44.6 ± 8.0; development subset n = 80, age=46.6 ± 7.1). We specifically evaluated three of the most promising of them - fluid-attenuated inversion recovery squared (FLAIR<sup>2</sup>), T1/T2 ratio, and artificial intelligence-derived double inversion recovery (AI-DIR) - and introduced a new combined contrast called multi-modal cortical lesion enhanced (MMCLE). We also harnessed transformer-based semantic segmentation to improve automated detection and delineation of these lesions. At baseline, we detected 14.8 + /-20.72 lesions per participant, with 86.0% true positive rate and 8.4% false positive rate across subjects for blinded MMCLE, using simultaneous review of all contrasts as the reference. High reproducibility was observed across field strengths and acquisition types (ICC 88.8-92.5%). We confirmed that cortical lesions can be clearly visualized and quantified with these methods. Using deep learning, we also confirmed that the simultaneous use of multiple contrasts improves quantification.