Year: 2023
Type: article
Abstract: Currently, the human race is facing several health-related issues where brain tumours are recognized as one of the leading causes of morbidity and mortality worldwide. Several researches and surveys have reported that the timely detection and prediction of brain tumours can help prevent their diverse impacts. Therefore, segmentation of these brain tumors is one of the prime tasks to proliferate the accuracy of diagnosis. Deep learning has become a promising technique to facilitate an automated Brain Tumor Segmentation (BTS) approach. Most current deep learning approaches rely on Convolutional Neural Networks (CNNs), which fail to contain long-term dependencies and global context information. Moreover, the performance of these systems is affected due to receptive field limitations during convolution operations. Currently, UNet-based architectures are adopted to perform medical image segmentation. Thus, in this work, we considered UNet as the base model for segmentation and incorporated transformer-based modules to improvise the segmentation accuracy; along with this, we present a hybrid attention mechanism that uses local and global context information. Based on this architecture, we evaluated the efficiency of the proposed approach for various Brats Datasets (2015, 2017, 2019, 2020 and 2021). The proposed approach achieves the average dice score of 0.94, 0.921, 0.83, and 0.94 for Brat's dataset.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4386936014