Comparison of post reconstruction- and reconstruction-based deep learning denoising methods in cardiac SPECT

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Myocardial perfusion SPECT (MPS) image quality is often compromised by low count statistics, potentially leading to diagnostic errors.
• Deep learning (DL) offers promising solutions for enhancing low-count MPS studies.

Purpose of the Study:

• To compare the effectiveness of post-reconstruction versus reconstruction-based DL denoising methods for MPS.
• To evaluate the impact of these DL methods on image noise, contrast, and defect detection.

Main Methods:

• A UNET-type DL network was trained on low-activity MPS data.
• Two DL approaches were evaluated: post-reconstruction denoising (OSEM+DL) and a DL penalty integrated into reconstruction (DLP).
• Performance was assessed using noise levels, myocardium-ventricle contrast, and signal-to-noise ratio (SNR) for artificial perfusion defects.

Main Results:

• OSEM+DL achieved the lowest noise levels across all activities.
• Both OSEM+DL and DLP demonstrated superior defect detection compared to standard OSEM.
• DLP provided higher SNR and better myocardium-ventricle contrast than OSEM+DL.

Conclusions:

• Both post-reconstruction and reconstruction-based DL denoising significantly improve MPS image quality.
• The choice between methods involves a trade-off between image smoothness and defect detection performance.