TransU²-Net: An Effective Medical Image Segmentation Framework Based on Transformer and U²-Net

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• U-Net and U²-Net architectures have advanced medical image segmentation.
• U²-Net, while effective, can overfit in medical image segmentation tasks due to excessive nesting.
• There is a need for efficient and accurate models for brain tumor segmentation.

Purpose of the Study:

• To propose TransU²-Net, a 2D network combining transformers and a lightweight U²-Net for automatic brain tumor segmentation in MRI.
• To enhance multi-scale feature extraction and global information capture for improved segmentation performance.

Main Methods:

• Developed a lightweight U²-Net architecture to capture multi-scale information and reduce redundant feature extraction.
• Integrated transformer blocks within stacked convolutional layers to capture global context.
• Employed skip-connections within the transformer to enhance spatial information representation.
• Introduced a novel multi-scale feature map fusion strategy for postprocessing.

Main Results:

• TransU²-Net achieved an average Dice coefficient of 88.17% on the BraTS2021 dataset for brain tumor segmentation.
• The model obtained a Dice coefficient of 74.69% on the MSD dataset for tumor evaluation.
• Comparative analysis demonstrated superior performance against existing 2D segmentation methods.

Conclusions:

• TransU²-Net offers an effective automatic medical image segmentation method by integrating transformers and U²-Net.
• The proposed method shows significant clinical importance and outperforms other 2D segmentation techniques.
• Experiments were conducted on publicly available datasets (BraTS2021, MSD) in accordance with ethical guidelines.