CSAM: A 2.5D Cross-Slice Attention Module for Anisotropic Volumetric Medical Image Segmentation.

Area of Science:

• Medical imaging
• Deep learning
• Computer vision

Background:

• Anisotropic medical data, common in MRI, presents challenges for 3D and 2D deep learning segmentation.
• Existing 3D methods struggle with anisotropic data, while 2D methods ignore volumetric context.

Purpose of the Study:

• To develop an efficient method for segmenting anisotropic volumetric medical data.
• To address limitations of existing 2D and 3D deep learning approaches.

Main Methods:

• Proposed a novel Cross-Slice Attention Module (CSAM) for 2.5D deep learning models.
• CSAM utilizes semantic, positional, and slice attention on multi-scale feature maps.
• The module is designed with minimal trainable parameters.

Main Results:

• Extensive experiments demonstrated the effectiveness and generalizability of CSAM.
• CSAM successfully captured cross-slice volumetric information for improved segmentation.
• The module showed strong performance across various network architectures and segmentation tasks.

Conclusions:

• CSAM offers an efficient and effective solution for segmenting anisotropic medical imaging data.
• The proposed module enhances deep learning segmentation by integrating volumetric context.
• CSAM represents a valuable advancement in medical image analysis techniques.