Towards lower-dose PET using physics-based uncertainty-aware multimodal learning with robustness to...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiological Physics

Background:

• Positron emission tomography (PET) imaging is limited by radiation exposure, especially in sensitive populations like pregnant women and children.
• Deep neural network (DNN) methods can enhance low-quality PET images but struggle with out-of-distribution (OOD) data.
• Current DNNs for PET-MRI image translation often fail to generalize to new data with different statistical properties.

Purpose of the Study:

• To develop a novel deep neural network (DNN) framework for improving low-dose/low-count PET image quality.
• To enhance the robustness of PET image reconstruction to out-of-distribution (OOD) acquisitions.
• To provide uncertainty quantification in PET image estimation using multimodal MRI data.

Main Methods:

• Proposed a sinogram-based uncertainty-aware DNN framework (suDNN) incorporating PET imaging physics.
• Modeled output uncertainty via per-voxel heteroscedasticity of residuals.
• Utilized multimodal input: low-dose PET images and corresponding multi-contrast MRI images.

Main Results:

• suDNN demonstrated improved robustness to OOD acquisitions compared to existing methods.
• Quantitative and qualitative results showed benefits of suDNN on in vivo simultaneous PET-MRI data.
• The framework successfully estimated standard-dose PET images from low-dose inputs with enhanced quality.

Conclusions:

• The suDNN framework offers a robust solution for high-quality PET imaging from low-dose acquisitions.
• This approach is particularly valuable for radiation-sensitive populations requiring reduced radiation exposure.
• The uncertainty-aware, physics-informed DNN improves PET image reconstruction reliability for diverse clinical scenarios.