Large-Kernel Attention for 3D Medical Image Segmentation.

Area of Science:

• Medical imaging analysis
• Deep learning for healthcare
• Computational anatomy

Background:

• Accurate segmentation of organs and tumors in 3D medical images (MRI, CT) is crucial for cancer diagnosis and treatment.
• Challenges include overlapping organs, anatomical variations, low contrast, diverse tumor characteristics, and background noise.
• Existing deep learning methods struggle with these complexities.

Purpose of the Study:

• To propose a novel 3D large-kernel (LK) attention module for enhanced 3D medical image segmentation.
• To improve the accuracy of multi-organ and tumor segmentation in challenging medical scans.
• To integrate the LK attention module into Convolutional Neural Networks (CNNs), specifically U-Net.

Main Methods:

• Developed a 3D large-kernel (LK) attention module combining self-attention and convolution.
• Incorporated local context, long-range dependencies, and channel adaptation.
• Decomposed LK convolution to optimize computational cost.
• Integrated the module into a U-Net architecture and evaluated on CT-ORG and BraTS 2020 datasets.

Main Results:

• The proposed 3D LK attention module significantly improved segmentation accuracy for organs and tumors.
• The best performing model, a Mid-type 3D LK attention-based U-Net, achieved state-of-the-art results.
• Performance gains were statistically validated against leading CNN and Transformer-based methods.
• Ablation experiments confirmed the effectiveness of convolutional decomposition and network design.

Conclusions:

• The novel 3D LK attention module effectively addresses challenges in 3D medical image segmentation.
• The proposed method achieves superior performance in multi-organ and tumor segmentation.
• This approach offers a promising advancement for automated medical image analysis in clinical settings.