Efficient brain tumor segmentation using Swin transformer and enhanced local self-attention.

Area of Science:

• Medical image analysis
• Artificial intelligence in neuro-oncology

Background:

• Convolutional Neural Networks (CNNs) have limitations in capturing long-range dependencies for brain tumor segmentation due to localized receptive fields.
• Vision Transformers (ViTs) leverage self-attention mechanisms to dynamically aggregate spatial information, outperforming CNNs in various segmentation tasks.

Purpose of the Study:

• To propose a novel network architecture for semantic brain tumor segmentation based on the Swin transformer.
• To address the limitations of CNNs in learning long-range dependencies by integrating transformer capabilities.

Main Methods:

• The proposed method employs an encoder-decoder structure combining Transformer and CNN modules.
• The encoder utilizes Enhanced Local Self-Attention (ELSA) transformer blocks for improved local feature extraction.
• Channel squeeze and spatial excitation blocks are incorporated to enhance feature representation.

Main Results:

• The approach was evaluated on the BraTS 2021 dataset, comprising 1251 cases with 3D MRI scans.
• The method achieved a high average Dice score of 89.77% and an average Hausdorff distance of 8.90 mm.
• The framework demonstrated excellent segmentation performance on multi-modal 3D brain MRI data.

Conclusions:

• An automated framework for brain tumor segmentation was developed using Swin transformer and enhanced local self-attention.
• The proposed method surpasses existing state-of-the-art 3D algorithms in brain tumor segmentation accuracy.
• The findings highlight the potential of transformer-based architectures for advanced medical image segmentation.