Enhanced Brain Stroke Lesion Segmentation in MRI Using a 2.5D Transformer Backbone U-Net Model.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Neuroimaging

Background:

• Accurate segmentation of brain stroke lesions from MRI is crucial for diagnosis and treatment planning.
• Existing methods face challenges in balancing accuracy and computational complexity.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for precise brain stroke lesion segmentation.
• To enhance the U-Net architecture with a transformer-based backbone and a 2.5D approach.

Main Methods:

• Implemented a U-Net model with a Mix Vision Transformer (MiT) backbone.
• Utilized a 2.5D method for processing 3D MRI data slices.
• Evaluated performance on the 2015 ISLES dataset using 4-fold cross-validation.

Main Results:

• The proposed U-Net with MiT backbone and 2.5D method achieved superior performance.
• Achieved Dice Coefficient of 0.8153 ± 0.0101 and IoU of 0.7835 ± 0.0079.
• Outperformed other state-of-the-art models including CNN-based UNet, nnU-Net, TransUNet, and SwinUNet.

Conclusions:

• Integrating transformer backbones and 2.5D techniques significantly advances brain stroke lesion segmentation.
• The developed model offers a more reliable and efficient tool for clinical diagnostic applications.