UNETR++: Delving Into Efficient and Accurate 3D Medical Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Transformer models show promise in 3D medical segmentation, but self-attention's quadratic complexity is a bottleneck.
• Volumetric medical imaging data presents computational challenges for existing transformer architectures.

Purpose of the Study:

• To develop an efficient and accurate 3D medical image segmentation approach named UNETR++.
• To address the computational limitations of self-attention mechanisms in transformer models for medical imaging.

Main Methods:

• Introduced a novel efficient paired attention (EPA) block with linearly complex spatial attention.
• EPA utilizes inter-dependent spatial and channel attention branches with shared weights for query and key mapping.
• Implemented UNETR++ architecture for 3D medical image segmentation.

Main Results:

• UNETR++ achieved state-of-the-art Dice Score of 87.2% on the Synapse benchmark.
• Significantly reduced parameters and FLOPs by over 71% compared to existing methods.
• Demonstrated effectiveness across five diverse medical imaging benchmarks (Synapse, BTCV, ACDC, BraTS, Decathlon-Lung).

Conclusions:

• UNETR++ offers a highly accurate and computationally efficient solution for 3D medical image segmentation.
• The proposed EPA block effectively captures discriminative spatial and channel-wise features with linear complexity.
• The approach shows strong potential for practical applications in medical image analysis.