Learned spatiotemporal correlation priors for CEST image denoising using incorporated global-spectral convolution...

Area of Science:

• Biomedical Imaging
• Artificial Intelligence in Medical Imaging
• Image Processing

Background:

• Chemical Exchange Saturation Transfer (CEST) imaging is crucial for various medical applications.
• Low signal-to-noise ratio (SNR) in CEST images presents a significant challenge for accurate analysis.
• Existing denoising methods often struggle with the complex noise patterns in CEST data.

Purpose of the Study:

• To develop a novel deep learning-based method, Denoising CEST Network (DECENT), for effective CEST image denoising.
• To exploit the inherent spatiotemporal correlation within CEST images for improved noise reduction.
• To provide a robust and efficient solution for enhancing CEST image quality in low SNR conditions.

Main Methods:

• DECENT utilizes a dual-pathway U-Net architecture with varying convolution kernel sizes to capture global and spectral features.
• A fusion pathway integrates features from both parallel pathways for comprehensive denoising.
• The method was validated using numerical simulations, phantom experiments, and in vivo studies on mouse brains and human skeletal muscle.

Main Results:

• DECENT demonstrated superior denoising performance compared to state-of-the-art methods like NLmCED, MLSVD, and BM4D, as evidenced by peak SNR (PSNR) and structural similarity index (SSIM) metrics.
• The network effectively reduced Rician noise in simulated and experimental low SNR CEST images.
• DECENT offers a computationally efficient alternative, avoiding complex parameter tuning and lengthy iterative processes.

Conclusions:

• DECENT successfully leverages the spatiotemporal correlation prior of CEST images to restore noise-free observations from noisy data.
• The proposed deep learning approach significantly outperforms existing denoising techniques.
• DECENT offers a promising tool for improving the diagnostic utility of CEST imaging in clinical and research settings.