Towards lower-dose PET using physics-based uncertainty-aware multimodal learning with robustness to out-of-distribution data

Radiation exposure in positron emission tomography (PET) imaging limits its usage the studies of radiation-sensitive populations, e.g., pregnant women, children, and adults that require longitudinal imaging. Reducing PET radiotracer dose or acquisition time reduces photon counts, which can deteriorate image quality. Recent deep-neural-network (DNN) based methods for image-to-image translation enable mapping low-quality images (acquired using substantially reduced dose), coupled with associated magnetic resonance (MRI) images, to high-quality images. However, such DNN focus on applications involving test data match statistical characteristics training very closely give little attention evaluating performance these DNNs new out-of-distribution (OOD) acquisitions. We propose a novel formulation models (i) underlying sinogram-based physics system (ii) uncertainty output through per-voxel heteroscedasticity residuals between predicted reference Our uncertainty-aware framework, namely, suDNN, estimates standard-dose multimodal input form low-dose/low-count corresponding multi-contrast MRI leading improved robustness suDNN OOD Results vivo simultaneous PET-MRI, various forms show benefits over current state art, quantitatively qualitatively.