Hybrid Channel Attention Regression U-Net (ARU-Net): An Enhanced Architecture for Brain Tumour Segmentation in Magnetic Resonance Imaging

Abstract

This research presents an accurate segmentation of brain tumours which is important in diagnosis and treatment planning of magnetic resource imaging (MRI) scans. Despite existing methodologies based on deep learning models, it remains difficult to distinguish between a tumour region and a normal brain tissue. To tackle this problem, a recent work proposed a new brain tumour segmentation technique on MRI, known as the Hybrid Attention Regression U-Net (ARU-Net) architecture. To enhance the effectiveness of segmentation, the proposed architecture utilises modern approaches including attention mechanisms, feature fusion processes, and regression-based output layer. It specifically uses Spatial and channel-wise attention mechanisms to capture important regions and inter channel dependencies. Spatial attention module in this architecture is based on improved Convolutional Block Attention Module which constrain attention on informative areas and suppress irrelevant or noisy ones. On the other hand, the channel-wise attention part, the Squeeze-and-Excitation (SE) blocks, recalibrates channel-wise feature responses for higher emphasis on salient features. The performance of proposed method has been evaluated comprehensively on the BraTS-2021 dataset where it shows enhanced accuracy over existing methods in detecting the tumour area. From the point of epoch curve, F1 score, the area under the curve, specificity, sensitivity, and accuracy in augmented and non-augmented datasets validated the stability of the proposed ARU-Net. This underlines its roles in segmentation of brain tumour, which if optimally developed holds the promise to significantly contribute for diagnosis and treatment planning.

Keywords

• Brain tumour imaging
• segmentation
• magnetic resonance imaging
• UNet
• convolutional neural network
• squeeze and excitation