A Dense U-Net Architecture for Multiple Sclerosis Lesion Segmentation

Abstract

Lesion segmentation is pivotal for the identification and cure of Multiple Sclerosis (MS). It is the need of the hour to highlight the neural networks with fewer layers of convolution, less training data samples while maintaining high accuracy and less processing speed to achieve real-time lesion segmentation. This paper evaluates a network with neural blocks comprising two jump points which are responsible for forward feeding of input data within the block. The downsampling path of the network architectures ensures transition invariance with reducing parameters while the upsampling path results in accurate localization. Data augmentation in the preprocessing part of the architecture helps achieving desired result with limited annotated data samples. We argue that the proposed net has numerous advantages: ReLu, relatively fewer layers and input output merging helps rectify vanishing gradient problem while decreasing parameters helps increase performance. The presented model has achieved promising results which signifies the efficacy in terms of some of the well known quality matrices. The dice score, sensitivity and specificity of the segmented lesions is computed on MICCAI MSSEG 2016 dataset. The method is entirely automated, efficient and fairly precise. Supplementary improvements in the model could provide a more robust and cost efficient approach to traditional medical image segmentation techniques.