Automated segmentation of multiple sclerosis lesions based on convolutional neural networks

Abstract

Automatic segmentation using Deep Learning (DL) has recently outperformed previous methods. In this paper, we are interested in automatic MS lesion segmentation of 2D images using DL approaches. The main challenge is to propose a new model that uses referenced CNN models such as U-Net, ResNet, and DenseNet, but with less parameters and a faster execution time. We evaluated the performance of our suggested technique, ‘Concat-U-Net’, to that of three implemented models, using two different datasets: our private MS database and the public ISBI. A comparison research was conducted using the Dice, Sensitivity, Precision, and PR-Curve evaluation metrics. On our private data set, the best Dice produced by the suggested technique was 0.80 as well as 0.77, 0.98, and 0.78 for sensitivity, PR-curve, and Precision, respectively. We enhanced MS segmentation results using the ISBI database and acquired a Dice value of 0.72, which is close to the results reported in the literature. For our data set and ISBI database, we lowered the elapsed execution time from 16.99 s to 9.36 s and from 25.02 s to 4.37 s, respectively. Furthermore, by reducing the number of parameters, we achieved an 81.51% time gain.