Comparison of deep learning-based denoising methods in cardiac SPECT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Myocardial perfusion SPECT (MPS) images are prone to noise artifacts from low-count statistics, potentially leading to misdiagnosis.
• Deep learning (DL) offers a promising solution for denoising MPS images and improving diagnostic accuracy.

Purpose of the Study:

• To compare the performance of various deep learning (DL) denoising models for myocardial perfusion SPECT (MPS) images.
• To evaluate the impact of DL denoising on image quality, noise levels, and defect detection.

Main Methods:

• Four DL models (CNN, RES, UNET, cGAN) were trained on MPS data reconstructed with varying acquisition times.
• Methods were compared against each other and standard reconstruction (OSEM) using noise metrics (CoV), image similarity (SSIM), and ROC analysis for defect detection.

Main Results:

• All DL methods significantly reduced noise compared to OSEM. CNN, RES, and UNET showed comparable defect detection performance to full-time acquisition, while cGAN had the lowest noise but also the poorest detection.
• DL denoising improved image quality and uniformity, though SSIM was higher with reduced acquisition time OSEM.

Conclusions:

• DL-based denoising enhances noise reduction in MPS imaging, with some models maintaining high diagnostic performance.
• Conditional GAN (cGAN) achieved the lowest noise levels but demonstrated the weakest defect detection among the evaluated DL methods.