Attention-TransBTS: A Novel Transformer-Unet Architecture for MRI Brain Tumor Segmentation

Abstract

This paper presents a novel model termed as Attention-TransBTS, designed for MRI brain tumor volumetric segmentation. The proposed model synergizes Transformer and CNN architectures, capitalizing on the comprehensive global modeling prowess of Transformer to mitigate the intrinsic limitations of CNN's local context comprehension. Simultaneously, it harnesses 3D CNN to extract spatial voxel features, thereby considerably preserving the spatial context intricacies embedded within MRI image data. Addressing the intricate challenge of uncertain tumor region localization and complex structural morphology within MRI images, we introduce a pioneering 3D CBAM attention module. This module systematically amplifies the feature extraction prowess of the segmentation network, specifically for irregular lesion domains, thereby meticulously achieving precision in the segmentation of tumor regions within MRI images. The 3D CBAM module, an evolutionary iteration of the CBAM module, adeptly expands its purview into three dimensions, effectively tailoring its applicability for 3D volumetric segmentation tasks. Experimental results on BraTS 2019 datasets demonstrate that our Attention-TransBTS outperforms the current state-of-the-art 3D methods for brain tumor segmentation on 3D MRI scans.