Motion correction of respiratory-gated PET images using deep learning based image registration framework

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Patient motion during PET scans causes artifacts, degrading image quality.
• Respiratory gating and non-rigid registration are used for correction but can be time-consuming.
• Deep learning offers a potential solution for efficient motion correction.

Purpose of the Study:

• To develop and validate an unsupervised deep learning framework for non-rigid image registration to correct motion artifacts in PET.
• To compare the performance of the deep learning method against iterative registration and ungated reconstructions.

Main Methods:

• An unsupervised deep learning network with a differentiable spatial transformer layer was designed for motion correction.
• Estimated deformation fields were integrated into an iterative image reconstruction algorithm.
• The method was validated using both simulated and clinical PET data.

Main Results:

• Deep learning motion correction significantly reduced normalized root mean square error (NRMS) in simulations (24.3%) compared to iterative registration (31.1%) and ungated images (41.9%).
• The method enhanced image details, reduced bias without increasing noise, and improved lesion contrast and liver boundaries in clinical data.
• AI-based correction outperformed iterative methods, offering higher contrast at matched noise levels.

Conclusions:

• Unsupervised deep learning-based non-rigid registration effectively corrects patient motion artifacts in PET imaging.
• This AI approach provides superior image quality and diagnostic information compared to conventional methods.
• The proposed framework holds promise for improving PET scan accuracy and efficiency.