Multi-attention Mechanism for Enhanced Pseudo-3D Prostate Zonal Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Prostate zonal segmentation in MRI is crucial for diagnosis and treatment planning.
• Existing methods often face challenges in balancing accuracy and computational efficiency.

Purpose of the Study:

• To introduce and evaluate a novel pseudo-3D Global-Local Channel Spatial Attention (GLCSA) mechanism.
• To enhance prostate zonal segmentation in high-resolution T2-weighted MRI images.

Main Methods:

• Developed a GLCSA mechanism integrating global and local attention in channel and spatial domains.
• Incorporated a slice interaction module to simulate 3D processing within a U-Net architecture.
• Evaluated on proprietary (44 patients) and public (ProstateX, 204 patients) datasets using Dice Similarity Coefficient (DSC) and Mean Surface Distance (MSD).

Main Results:

• GLCSA significantly improved segmentation accuracy for transition zone (TZ) and peripheral zone (PZ) with minimal parameter increase (1.27%).
• Achieved substantial DSC increases on both datasets, notably 7.34% (TZ) and 24.80% (PZ) on ProstateX.
• GLCSA-UNet demonstrated competitive performance against 2D, 2.5D, and 3D models.

Conclusions:

• GLCSA provides a robust balance between the detailed feature extraction of 3D models and the efficiency of 2D models.
• The proposed mechanism is a promising tool for improving prostate MRI image segmentation accuracy.
• Ablation studies confirmed the effectiveness of combined attention mechanisms and global embedding.