In this paper, we present a fully convolutional densely connected network (Tiramisu) for multiple sclerosis (MS) lesion segmentation. Different from existing methods, we use stacked slices from all three anatomical planes to achieve a 2.5D method. Individual slices from a given orientation provide global context along the plane and the stack of adjacent slices adds local context. By taking stacked data from three orientations, the network has access to more samples for training and can make more accurate segmentation by combining information of different forms. The conducted experiments demonstrated the competitive performance of our method. For an ablation study, we simulated lesions on healthy controls to generate images with ground truth lesion masks. This experiment confirmed that the use of 2.5D patches, stacked data and the Tiramisu model improve the MS lesion segmentation performance. In addition, we evaluated our approach on the Longitudinal MS Lesion Segmentation Challenge. The overall score of 93.1 places the L 2-loss variant of our method in the first position on the leaderboard, while the focal-loss variant has obtained the best Dice coefficient and lesion-wise true positive rate with 69.3% and 60.2%, respectively.
Keywords: Deep learning; Multiple sclerosis; Segmentation.