Volumetric Segmentation of Brain Tumors Using Neural Networks
Ștefana Duță, Alina Elena Sultana · 2022 E-Health and Bioengineering Conference (EHB) · 2022
Although medical images can provide exceptional anatomical views, distinguishing the pixels of lesions from those of healthy tissue remains one of the most challenging tasks in medical image analysis. This paper focuses on the task of automatic brain tumors segmentation from Magnetic Resonance Imaging (MRI) images. The aim of the study is to identify an optimal solution based on volumetric convolutional neural architectures, trained on the Brain Tumor Segmentation (BraTS) 2020 dataset. This work presents some of the most popular Convolutional Neural Networks-like volumetric neural networks (Seg-Net 3D, U-Net 3D, ResUNet) used for automatic segmentation of brain tumors in a comparative manner and how they can be optimized to achieve maximum performance. The best Dice score results are obtained after training the ResUNet type network, which leads to a percentage of 91.4% for the whole tumor, while the U-Net network reaches a maximum value of Dice score of 90.8%. The worst results are obtained by using the SegNet architecture, that leads to a dice score of 90.4%.