Abstract
In conventional non-quantitative magnetic resonance imaging, image contrast is consistent within images, but absolute intensity can vary arbitrarily between scans. For quantitative analysis of intensity data, images are typically normalized to a consistent reference. The most convenient reference is a tissue that is always present in the image, and is unlikely to be affected by pathological processes. In multiple sclerosis neuroimaging, both the white and gray matter are affected, so normalization techniques that depend on brain tissue may introduce bias or remove biological changes of interest. We introduce a complementary procedure, image "calibration," the goal of which is to remove technical intensity artifacts while preserving biological differences. We demonstrate a deep learning approach to segmenting fat from within the orbit of the eyes on T
-weighted images at 1.5 and 3 T to use as a reference tissue, and use it to calibrate 1018 scans from 256 participants in a study of pediatric-onset multiple sclerosis. The machine segmentations agreed with the adjudicating expert (DF) segmentations better than did those of other expert humans, and calibration resulted in better agreement with semi-quantitative magnetization transfer ratio imaging than did normalization with the WhiteStripe
algorithm. We suggest that our method addresses two key priorities in the field: (1) it provides a robust option for serial calibration of conventional scans, allowing comparison of disease change in persons imaged at multiple time points in their disease; and (ii) the technique is fast, as the deep learning segmentation takes only 0.5 s/scan, which is feasible for both large and small datasets.