Cross-Attention for Improved Motion Correction in Brain PET.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Head movement during positron emission tomography (PET) scans causes artifacts, degrading image quality and limiting clinical diagnosis.
• Existing deep learning methods for PET motion correction require hardware motion tracking and struggle with cross-subject generalizability.

Purpose of the Study:

• To develop a deep learning model integrating a cross-attention mechanism for robust head motion correction in PET imaging.
• To improve the generalizability of motion correction across different subjects and PET scanners without hardware motion tracking.

Main Methods:

• A supervised deep learning network incorporating a cross-attention mechanism was developed to learn head motion from PET data.
• Cross-attention identifies spatial correspondences between reference and moving images, focusing on the head region for correction.
• The model was validated on brain PET data from two scanners (HRRT and mCT) with varying capabilities (ToF).

Main Results:

• The cross-attention model significantly improved motion correction performance compared to traditional and deep learning benchmarks.
• In HRRT studies, improvements of 58% (translation) and 26% (rotation) were observed across multiple subjects.
• In mCT studies, performance enhancements of 66% (translation) and 64% (rotation) were achieved.

Conclusions:

• The proposed cross-attention mechanism enhances the robustness and accuracy of deep learning-based PET motion correction.
• This approach shows potential for improving brain PET image reconstruction quality without reliance on hardware motion tracking.
• The method offers a promising solution for more reliable clinical diagnosis and treatment planning using PET imaging.