Transformer-augmented lightweight U-Net (UAAC-Net) for accurate MRI brain tumor segmentation

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Neuro-oncology

Background:

• Accurate brain tumor segmentation in MRI is vital for clinical decision-making.
• Glioma segmentation requires sophisticated models to capture complex tumor characteristics.
• Existing segmentation methods face challenges in handling diverse tumor structures and image variations.

Purpose of the Study:

• To develop a novel, lightweight, transformer-based U-Net model for enhanced brain tumor segmentation.
• To improve the capture of long-range dependencies and contextual information in MRI scans.
• To achieve superior performance in segmenting gliomas compared to current state-of-the-art techniques.

Main Methods:

• Implementation of a transformer-based U-Net architecture incorporating attention mechanisms.
• Utilizing atrous convolution for multi-layer feature extraction and capturing global context.
• Evaluation on the BraTS 2020 dataset using Dice coefficient, accuracy, mean IoU, sensitivity, and specificity.

Main Results:

• The proposed model demonstrated superior performance against established methods like MimicNet and Swin Transformer-based UNet.
• Achieved high segmentation accuracy (98.23%), Dice score (0.9716), and mean IoU (0.8242) on the BraTS 2020 dataset.
• Effectively addressed various segmentation challenges, indicating robustness in clinical applications.

Conclusions:

• The developed lightweight, transformer-based U-Net model offers a significant advancement in brain tumor segmentation accuracy.
• The model's ability to integrate attention and multi-layer feature extraction enhances its capacity for precise glioma segmentation.
• This approach holds promise for improving diagnostic accuracy and treatment monitoring in neuro-oncology.