A generalizable diffusion framework for 3D low-dose and few-view cardiac SPECT imaging.

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence

Background:

• Myocardial perfusion imaging using SPECT is crucial for diagnosing coronary artery disease.
• Low-dose and few-view SPECT acquisitions can compromise image quality, limiting diagnostic accuracy.
• Existing deep learning methods struggle to generalize across different acquisition settings.

Purpose of the Study:

• To develop a robust deep learning framework, DiffSPECT-3D, for improving 3D cardiac SPECT image quality.
• To enable effective SPECT imaging in low-dose and few-view settings without compromising clinical performance.
• To ensure the framework generalizes across diverse acquisition parameters without retraining.

Main Methods:

• Introduced DiffSPECT-3D, a diffusion framework for 3D cardiac SPECT imaging.
• Implemented a consistency strategy using image and projection data for alignment with measurements and scanner geometry.
• Utilized a 2.5D conditional strategy incorporating CT anatomical information and total variation for 3D contextual observation.

Main Results:

• Extensive evaluation on 1,325 clinical SPECT studies across varied low-dose (1%-50%) and few-view (1-9 views) settings.
• Demonstrated generalization across different low-dose and few-view acquisition parameters without network retraining.
• Achieved image quality comparable to full-dose SPECT, validated against cardiac catheterization and expert review.

Conclusions:

• DiffSPECT-3D offers a promising solution for high-quality low-dose and few-view cardiac SPECT imaging.
• The framework shows potential for improving image quality in standard and stress-first cardiac SPECT protocols.
• The method generalizes effectively, addressing limitations of previous deep learning approaches in SPECT imaging.