Automatic segmentation of the spinal cord and intramedullary multiple\n sclerosis lesions with convolutional neural networks

The spinal cord is frequently affected by atrophy and/or lesions in multiple\nsclerosis (MS) patients. Segmentation of the spinal cord and lesions from MRI\ndata provides measures of damage, which are key criteria for the diagnosis,\nprognosis, and longitudinal monitoring in MS. Automating this operation\neliminates inter-rater variability and increases the efficiency of\nlarge-throughput analysis pipelines. Robust and reliable segmentation across\nmulti-site spinal cord data is challenging because of the large variability\nrelated to acquisition parameters and image artifacts. The goal of this study\nwas to develop a fully-automatic framework, robust to variability in both image\nparameters and clinical condition, for segmentation of the spinal cord and\nintramedullary MS lesions from conventional MRI data. Scans of 1,042 subjects\n(459 healthy controls, 471 MS patients, and 112 with other spinal pathologies)\nwere included in this multi-site study (n=30). Data spanned three contrasts\n(T1-, T2-, and T2*-weighted) for a total of 1,943 volumes. The proposed cord\nand lesion automatic segmentation approach is based on a sequence of two\nConvolutional Neural Networks (CNNs). To deal with the very small proportion of\nspinal cord and/or lesion voxels compared to the rest of the volume, a first\nCNN with 2D dilated convolutions detects the spinal cord centerline, followed\nby a second CNN with 3D convolutions that segments the spinal cord and/or\nlesions. When compared against manual segmentation, our CNN-based approach\nshowed a median Dice of 95% vs. 88% for PropSeg, a state-of-the-art spinal cord\nsegmentation method. Regarding lesion segmentation on MS data, our framework\nprovided a Dice of 60%, a relative volume difference of -15%, and a lesion-wise\ndetection sensitivity and precision of 83% and 77%, respectively. The proposed\nframework is open-source and readily available in the Spinal Cord Toolbox.\n