Brain MRI study for glioma segmentation using convolutional neural networks and original post-processing techniques with low computational demand

TL;DR

This paper presents a glioma segmentation method using CNNs and novel low-demand post-processing techniques, achieving state-of-the-art accuracy on the BraTS 2018 dataset.

Contribution

It introduces a new segmentation approach combining CNNs with original low-complexity post-processing, improving accuracy without high computational costs.

Findings

Achieved Dice coefficients of 0.8934, 0.8376, and 0.8113 for different tumor regions.

Matched state-of-the-art results in glioma segmentation.

Utilized a low computational demand methodology.

Abstract

Gliomas are brain tumors composed of different highly heterogeneous histological subregions. Image analysis techniques to identify relevant tumor substructures have high potential for improving patient diagnosis, treatment and prognosis. However, due to the high heterogeneity of gliomas, the segmentation task is currently a major challenge in the field of medical image analysis. In the present work, the database of the Brain Tumor Segmentation (BraTS) Challenge 2018, composed of multimodal MRI scans of gliomas, was studied. A segmentation methodology based on the design and application of convolutional neural networks (CNNs) combined with original post-processing techniques with low computational demand was proposed. The post-processing techniques were the main responsible for the results obtained in the segmentations. The segmented regions were the whole tumor, the tumor core, and the enhancing tumor core, obtaining averaged Dice coefficients equal to 0.8934, 0.8376, and 0.8113, respectively. These results reached the state of the art in glioma segmentation determined by the winners of the challenge.