Brain tumor image segmentation based on Semantic Flow Guided Sampling and Attention Mechanism

The automatic segmentation method for brain tumors based on a U-shaped network structure of⁃ ten suffers from information loss due to multiple convolution and sampling operations，resulting in subopti⁃ mal segmentation results. To address this issue，this study proposed a feature alignment unit that utilizes semantic information flow to guide the up-sampling feature recovery and design designed a lightweight Du⁃ al Attention Feature Alignment Network（DAFANet）based on this unit. Firstly，to validate its effective⁃ ness and generalization，the feature alignment unit was introduced separately into three classic networks，namely 3D UNet，DMFNet，and HDCNet. Secondly，a lightweight dual-attention feature alignment net⁃ work named DAFANet was proposed based on DMFNet. The feature alignment unit enhanced feature restoration in the up-sampling process，and a 3D Expectation-Maximization attention mechanism was ap⁃ plied to both the feature alignment path and cascade path to capture the full contextual dependency. The generalized Dice loss function was also used to improve segmentation accuracy in the case of data imbal⁃ ance and accelerate model convergence. Finally，the proposed algorithm is validated on the BraTS2018 and BraTS2019 public datasets，achieving segmentation accuracies of 80. 44%，90. 07%，84. 57% and 78. 11%，90. 10%，82. 21% in the ET，WT，and TC regions，respectively. Compared to current popu⁃ lar segmentation networks，the proposed algorithm demonstrates better segmentation performance in en⁃ hancing tumor regions and is more adept at handling details and edge information. © 2024 Chinese Academy of Sciences. All rights reserved.