Multimodal MRI Segmentation of Brain Tissue and T2-Hyperintense White Matter Lesions in Multiple Sclerosis using Deep Convolutional Neural Networks and a Large Multi-center Image Database

Multiple sclerosis (MS) is a demyelinating disease that affects the central nervous system (CNS) and is characterized by the presence of CNS lesions. Volumetric measures of tissues, including lesions, on magnetic resonance imaging (MRI) play key roles in the clinical management and treatment evaluation of MS patient. Recent advances in deep learning (DL) show promising results for automated medical image segmentation. In this work, we used deep convolutional neural networks (CNNs) for brain tissue classification on MRI acquired from MS patients in a large multi-center clinical trial. Multi-channel MRI data that included T1-weighted, dual-echo fast spin echo, and fluid-attenuated inversion recovery images were acquired on these patients. The pre-processed images (following co-registration, skull stripping, bias field correction, intensity normalization, and de-noising) served as the input to the CNN for tissue classification. The network was trained using expert-validated segmentation. Quantitative assessment showed high Dice similarity coefficients between the CNN and the validated segmentation, with DSC values of 0.94 for white matter and grey matter, 0.97 for cerebrospinal fluid, and 0.85 for T2 hyperintense lesions. These results suggest that deep neural networks can successfully segment brain tissues, which is crucial for reliable assessment of tissue volumes in MS.