Segmentation of White Matter Lesions in MRI Images Using Optimization-Based Deep Neural Network

Abstract

White matter lesions (WML) cause numerous brain diseases, and automatic WML segmentation is essential for evaluating the natural disease course and the efficacy of clinical interventions, especially drug discovery. This work presents a novel metaheuristic Harris hawks’ optimization (HHO) to train the deep convolutional neural network (DCNN) as a robust model for WML segmentation. The proposed method was tested on MICCAI 2019 dataset. However, the WML regions would appear blurry and global inconsistency. In order to solve the fuzzy problem, we improved the network architecture by incorporating skip connections across the mirrored layers within encoder and decoder stacks. As a result, the segmented WML output appears more realistic and coherent with its surrounding contexts. The proposed optimization method, skip connections, has increased the segmentation performance of Dice score to 79.92%, precision 82.25%, recall 83.90%, and F1-score 82.94% for MRI dataset which is superior compared to other approaches. The proposed technique outperformed existing state-of-the-art methodologies in both qualitative and quantitative measurements across a wide range of medical modalities.