Higher SNR PET image prediction using a deep learning model and MRI image.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Positron Emission Tomography (PET) images often exhibit a poor signal-to-noise ratio (SNR), limiting diagnostic accuracy.
• Deep neural networks (DNNs) show promise for image denoising, but training often requires high-quality reference images.

Purpose of the Study:

• To develop and evaluate a DNN model that improves PET image SNR using co-registered MRI data.
• To achieve noise reduction and accelerated image reconstruction without relying on higher SNR PET training data.

Main Methods:

• A deep neural network (DNN) model comprising three modified U-Nets (3U-net) was proposed.
• Training involved PET images reconstructed with filtered-backprojection (FBP) as input and maximum likelihood expectation maximization (MLEM) as target.
• Digital brain phantoms with simulated Poisson noise and attenuation effects were used for evaluation, with additional noise introduced to training inputs.

Main Results:

• The proposed 3U-net model, trained with PET/MRI data, reduced mean squared error (MSE) by 34.0% compared to a 1U-net model trained with only PET data.
• This MSE reduction is equivalent to a 2.9-fold increase in PET data count levels.
• Lesion contrast-to-noise ratio (CNR) improved 2.7-fold (1U-net) and 1.4-fold (3U-net) compared to MLEM images, demonstrating enhanced lesion detectability.

Conclusions:

• The proposed DNN method effectively improves PET image SNR and accelerates reconstruction using readily available PET and MRI data.
• This approach offers a viable solution for enhancing PET image quality in clinical settings without the need for specialized high-SNR acquisitions.