Attention-based convolutional neural network for perfusion T2-weighted MR images preprocessing
This work addresses skull-stripping for perfusion MRI analysis in medical imaging, particularly for cases with abnormal brain anatomy, representing an incremental improvement over existing methods.
The researchers tackled the problem of skull-stripping in perfusion T2-weighted MR images, which is crucial for accurate perfusion analysis, by integrating spatial and channel squeeze-and-excitation attention mechanisms into a U-Net+ResNet architecture. They found that the scSE-POST integration strategy achieved the best result with an average Dice Coefficient of 0.9810.
Accurate skull-stripping is crucial preprocessing in dynamic susceptibility contrast-enhanced perfusion magnetic resonance data analysis. The presence of non-brain tissues impacts the perfusion parameters assessment. In this study, we propose different integration strategies for the spatial and channel squeeze and excitation attention mechanism into the baseline U-Net+ResNet neural network architecture to provide automatic skull-striping i.e., Standard scSE, scSE-PRE, scSE-POST, and scSE Identity strategies of plugging of scSE block into the ResNet backbone. We comprehensively investigate the performance of skull-stripping in T2-star weighted MR images with abnormal brain anatomy. The comparison that utilizing any of the proposed strategies provides the robustness of skull-stripping. However, the scSE-POST integration strategy provides the best result with an average Dice Coefficient of 0.9810.