Dynamic Linear Transformer for 3D Biomedical Image Segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Transformer neural networks excel in biomedical image segmentation via self-attention.
• Existing 2D methods overlook crucial 3D volume information.
• 3D Transformers face computational challenges due to quadratic complexity.

Purpose of the Study:

• To develop a 3D Transformer-based segmentation method with linear complexity.
• To address the limitations of current 2D and computationally intensive 3D segmentation approaches.
• To introduce a dynamic token concept for efficient self-attention calculation.

Main Methods:

• Proposed a novel encoder-decoder Transformer architecture with linear complexity.
• Introduced a dynamic token concept to reduce self-attention computation.
• Leveraged global information modeling for uncertainty map generation across hierarchy stages.

Main Results:

• Evaluated on challenging CT pancreas segmentation datasets.
• Demonstrated promising and highly feasible segmentation performance.
• Achieved accurate uncertainty quantification using single annotations.

Conclusions:

• The novel 3D Transformer effectively segments medical images while managing computational complexity.
• The method provides accurate uncertainty quantification, valuable for clinical applications.
• This work advances 3D medical image segmentation using efficient Transformer architectures.