Brain Tumor Semantic Segmentation using Residual U-Net++ Encoder-Decoder Architecture

Abstract

Image segmentation is considered one of the essential tasks for extracting useful information from an image. Given the brain tumor and its consumption of medical resources, the development of a deep learning method for MRI to segment the brain tumor of patients’ MRI is illustrated here. Brain tumor segmentation technique is crucial in detecting and treating MRI brain tumors. Furthermore, it assists physicians in locating and measuring tumors and developing treatment and rehabilitation programs. The residual U-Net++ encoder-decoder-based architecture is designed as the primary network, and it is an architecture that is hybridized between ResU-Net and U-Net++. The proposed Residual U-Net++ is applied to MRI brain images for the most recent and well-known global benchmark challenges: BraTS 2017, BraTS 2019, and BraTS 2021. The proposed approach is evaluated based on brain tumor MRI images. The results with the BraST 2021 dataset with a dice similarity coefficient (DSC) is 90.3%, sensitivity is 96%, specificity is 99%, and 95% Hausdorff distance (HD) is 9.9. With the BraST 2019 dataset, a DSC is 89.2%, sensitivity is 96%, specificity is 99%, and HD is 10.2. With the BraST 2017 dataset, a DSC is 87.6%, sensitivity is 94%, specificity is 99%, and HD is 11.2. Furthermore, Residual U-Net++ outperforms the standard brain tumor segmentation approaches. The experimental results indicated that the proposed method is promising and can provide better segmentation than the standard U-Net. The segmentation improvement could help radiologists increase their radiologist segmentation accuracy and save time by 3%.