Brain Tumor Segmentation from 3D MRI Scans Using U-Net

Abstract

A fully automated system based on three-dimensional (3D) magnetic resonance imaging (MRI) scans for brain tumor segmentation could be a diagnostic aid to clinical specialists, as manual segmentation is challenging, arduous, tedious and error prone. Employing 3D convolutions requires large computational cost and memory capacity. This study proposes a fully automated approach using 2D U-net architecture on BraTS2020 dataset to extract tumor regions from healthy tissue. All the MRI sequences are experimented with the model to determine for which sequence optimal performance is achieved. After normalization and rescaling, using optimizer Adam with learning rate 0.001 on T1 MRI sequence, we get an accuracy of 99.41% and dice similarity coefficient (DSC) of 93%, demonstrating the effectiveness of our approach. The model is further trained with different hyper-parameters to assess the robustness and performance consistency.