Abstract
Multiple sclerosis (MS) is characterized by central nervous system lesions detectable via MRI. Existing diagnostic criteria incorporate presence of white matter lesions, but specificity can be improved using MS-specific imaging biomarkers, including paramagnetic rim lesions (PRLs) and central vein sign (CVS). However, manual segmentation of lesions, PRLs, and CVS is time-consuming and subjective. We propose a fully-automated joint segmentation method called Automated Lesion, PRL, and CVS Analysis (ALPaCA). We trained ALPaCA using subject-level cross-validation on 47 adults with MS and 50 adults with radiological MS mimics. ALPaCA uses a voxel-wise lesion segmentation method to propose a large set of lesion candidates. Lesion candidates are input into a multi-contrast, multi-label 3D convolutional neural network as 3D patches to produce lesion, PRL, and CVS predictions. When multiple lesions exist within a patch, an attention mechanism identifies which lesion candidate to classify. At the lesion level, ALPaCA achieves cross-validation areas under the receiver operating characteristic curve (AUROCs) of 0.95, 0.91, and 0.87 for lesion, PRL, and CVS classification, outperforming previous methods (all p < 0.001). Correlations between subject-level ALPaCA lesion and PRL scores with manual counts are higher than those of previous methods (p < 0.001; p = 0.03). Subject-level ALPaCA PRL and CVS scores are highly associated with MS in logistic regressions, when controlling for age and sex (p < 0.001). ALPaCA allows for fully-automated simultaneous segmentation of MS lesions, PRLs, and CVS using clinically-feasible scans. These segmentations outperform existing methods at the lesion and subject level.