Uncertainty Assessment in Whole-Body Low Dose Pet Reconstruction Using Non-Parametric Bayesian Deep Learning Approach

The measurement of uncertainty in deep learning (DL) based reconstruction of Positron Emission Tomography (PET) images is imperative for its successful application in clinical settings, as it is vital for making informed clinical decisions. To address this issue, we propose a fully non-parametric Bayesian framework for uncertainty estimation in DL-based Ultra Low Dose PET image reconstruction. Our framework combines an Adam optimizer with stochastic gradient Langevin dynamics (SGLD) to sample from the underlying posterior distribution. The Bayesian Low Dose (LD) PET reconstruction approach shows a better correlation of the predicted uncertainty to the Dose Reduction Factor (DRF), as compared to the benchmark Monte-Carlo Dropout 95 percentile score (r²=0.9174 vs. r²=0.6144 on 10, 631 examples at the inference phase). Furthermore, statistically significant improvement was achieved in SSIM, PSNR and NRMSE measures for DRF's 4, 10, 20, 50 and 100 (paired t-test with p < 0.01).