Deep learning with noise-to-noise training for denoising in SPECT myocardial perfusion imaging

Area of Science:

• Medical imaging
• Nuclear medicine
• Artificial intelligence in healthcare

Background:

• Myocardial perfusion imaging (MPI) using single-photon emission computed tomography (SPECT) often requires post-reconstruction filtering for noise suppression due to limited data counts.
• Traditional filtering methods can impact image resolution and diagnostic accuracy.

Purpose of the Study:

• To investigate a deep learning (DL) approach for denoising SPECT-MPI images.
• To compare the effectiveness of DL denoising against traditional postfiltering in improving perfusion defect detectability.

Main Methods:

• A noise-to-noise (N2N) training approach was employed for denoising SPECT-MPI images, utilizing a coupled U-Net (CU-Net) architecture.
• Bootstrap procedure generated multiple noise realizations from clinical list-mode acquisitions for network training.
• Evaluated detection performance using non-prewhitening matched filter (NPWMF), assessed left ventricular (LV) wall uniformity, and quantified spatial resolution (FWHM).

Main Results:

• DL denoised images with CU-Net significantly improved perfusion defect detection compared to OSEM with Gaussian postfiltering across all contrast levels.
• Signal-to-noise ratio (SNR_D) in NPWMF output increased by 8% (P < 10^-4) over optimal Gaussian smoothing, with reduced inter-subject variability.
• CU-Net outperformed 3D nonlocal means (NLM) and convolutional autoencoder (CAE) denoising networks, and improved detection with less post-reconstruction smoothing (23% SNR_D increase).

Conclusions:

• The proposed DL approach with N2N training provides superior noise suppression in SPECT-MPI images compared to conventional postfiltering.
• DL with CU-Net demonstrates superior performance in perfusion defect detection over conventional 3D Gaussian filtering, NLM, and CAE methods.