A brain tumor segmentation method based on attention mechanism

Abstract

The rise in brain tumor incidence due to the global population aging has intensified the need for precise segmentation methods in clinical settings. Current segmentation networks often fail to capture comprehensive contextual information and fine edge details of brain tumors, which are crucial for accurate diagnosis and treatment. To address these challenges, we introduce BSAU-Net, a novel segmentation algorithm that employs attention mechanisms and edge feature extraction modules to enhance performance. Our approach aims to assist clinicians in making more accurate diagnostic and therapeutic decisions. BSAU-Net incorporates an edge feature extraction module (EA) based on the Sobel operator, enhancing the model's sensitivity to tumor regions while preserving edge contours. Additionally, a spatial attention module (SPA) is introduced to establish global feature correlations, critical for accurate tumor segmentation. To address class imbalance, which can hinder performance, we propose BADLoss, a loss function tailored to mitigate this issue. Experimental results on the BraTS2018 and BraTS2021 datasets demonstrate the effectiveness of BSAU-Net, achieving average Dice coefficients of 0.7506 and 0.7556, PPV of 0.7863 and 0.7843, sensitivity of 0.8998 and 0.9017, and HD95 of 2.1701 and 2.1543, respectively. These results highlight BSAU-Net's potential to significantly improve brain tumor segmentation in clinical practice.

Keywords: Attention mechanism; Brain tumor segmentation; Edge refinement.

Fig. 1

Overall framework of the proposed BSAU-Net.

Fig. 2

Structure of the EA module.

Fig. 3

Structure of the SPA module.

Fig. 4

Comparison of MRI image preprocessing effects. (a) The original image; (b) The image after initial processing; (c) The image after CLAHE enhancement.

Fig. 5

Comparative segmentation results of four modalities in a sample from the BraTS2018 dataset. From left to right: sample image, ground truth (mask), U-Net, U-Net++, MDU-net, AMSU-net and BSAU-net.

Fig. 6

Comparative segmentation results of four modalities in a sample from the BraTS2021 dataset. From left to right: sample image, ground truth (mask), U-Net, U-Net++, MDU-net, AMSU-net and BSAU-net.

Fig. 7

Comparison of detail regions in sample segmentation results of BraTS2018 dataset. From left to right: sample image, ground truth (mask), U-Net, U-Net++, MDU-net, AMSU-net and BSAU-net.

Fig. 8

Comparison of detail regions in sample segmentation results of BraTS2021 dataset. From left to right: sample image, ground truth (mask), U-Net, U-Net++, MDU-net, AMSU-net and BSAU-net.

Fig. 9

Segmentation mask refined by CRF. From left to right: original image, mask, processed mask.