2D FLAIR, 2D T1-weighted spin echo, 2D T2-weighted spin echo, and 2D PD-weighted images were collected from 455 MRI exams coming from 63 different scanners. Gold standard T2 lesion and T1 hypointensity masks were generated by two well-calibrated MD neuroimagers and reviewed by a senior MD neuroimager. A semantic segmentation network, SegResNet, was trained using all four image types and lesion masks as input with the goal of learning to simultaneously segment T2 lesions and T1 hypointensities without any user input. The trained network was evaluated on 75 additional MRI exams and resulting segmentations were compared to gold standard masks in terms of volumetric correlation and spatial overlap (Dice similarity coefficient).

Median (interquartile range) of the gold standard T2 lesion and T1 hypointensity volumes were 14.7 (5.7 – 29.8) and 3.2 (1.3 – 9.1) mL, respectively. Automatically segmented T2 lesion volume correlated with the corresponding gold standard volume at r=0.979, p <0.001 with an average Dice value of 0.69. Automatically segmented T1 hypointensity volume correlated with the corresponding gold standard volume at r=0.968, p<0.001, with an average Dice value of 0.55. Segmentation took on average 2 seconds per case.