Noise Adaptive Deep Convolutional Neural Network for Whole-Body PET Denoising

Many factors can impact PET image noise levels including injected dose, wait time, patient’s BMI, scan duration etc.. These factors can lead to a large spatial variation and inter-patient variation of PET image noise. Commonly used PET noise suppression methods include regularized image reconstruction and non-linear post-filtering, both of which require user selected functions and parameters to achieve a desired signal-to-noise tradeoff. This challenging task is further complicated by the intra-patient and inter-patient noise level variations. In this study, we investigated a deep convolutional neural network (DCNN) approach for whole-body PET image denoising that can adapt to different input noise levels to yield a consistent low noise PET image. We adopted an 8-layer residual network architecture that is trained with multiple noise samples paired with a single low noise high count PET image (label). All these noisy samples share the same clean latent image but are corrupted with different levels of noise. Thus, the network parameters are optimized to estimate the noise residual from input images with different noise levels. The training dataset constituted of 8 patient studies and 2 phantom studies. We evaluated the same DCNN architecture trained with multiple noise levels (DCNN-M) against the one trained with a single noise level (DCNN-S) on 2 real patient studies inserted with GATE simulated lesions. The results show that DCNN-M yielded superior performance in lesion contrast recovery compared to a Gaussian post-filter at matched liver CoV levels, and yielded more consistent coefficient of variation (CoV) in the liver across different scan durations than DCNN-S.