Research on brain tumor segmentation algorithm based on attention mechanism

Magnetic Resonance Imaging (MRI) playing a crucial role in the task of brain tumor segmentation. Achieving high-precision segmentation in brain tumor images is highly challenging. Currently, brain tumor segmentation mainly relies on manual segmentation by skilled medical professionals, which is time-consuming.In this paper, we propose a brain tumor segmentation algorithm, AUFP (Attention-based U-Net with Feature Pyramid), aimed at improving segmentation accuracy. AUFP introduces an Attention Feature Pyramid module (AFP) before the up-sampling stage in the U-Net architecture and an Attention Up-Sample module (AU) during the up-sampling process. The AFP module incorporates attention mechanisms to use global information effectively, enhancing segmentation accuracy.Experimental evaluations on the BraTS2017 data set indicate the effectiveness of the proposed AUFP algorithm. The overall tumor segmentation, core tumor segmentation, and enhancing tumor segmentation accuracy on the BraTS2017 data set are $\mathbf{9 2. 1 \%}, \mathbf{7 4. 4 \%}$, and $\mathbf{7 1. 1} \%$. Compared to a modified U-Net model (DUNet), with respective increases of $2.1 \%, 3.7 \%$, and $7.7 \%$. In conclusion, the introduced feature pyramid and attention up-sample modules in this paper effectively enhance brain tumor segmentation accuracy by improving the utilization of local information. Our results validate the efficacy of these modules for the task of brain tumor segmentation.