Deep learning convolutional neural networks (DCNN) have been successfully applied to suppress image noise in PET applications. Training a DCNN denoising model usually requires large amounts of paired datasets including high quality targets. In practice, obtaining high quality training targets is often very challenging due to limitations on radiation exposure and imaging times, in addition to the complications of voluntary (e.g. body) and involuntary motions (e.g. respiratory and muscle relaxation). In this study, inspired by the previous work of Noise2Noise training, we hypothesize that instead of training the network to map noisy samples to high quality targets, we could train a similar neural network by pairing one noise realization to an ensemble of noise realizations. We compared a deep residual network trained with three training schemes including one using high count target in training (HC target), one using noise to noise (N2N) training that maps one noise realization to another noise realization at the same count level, and the proposed noise to noise ensemble (N2NEN) training that maps one noise realization to multiple noise realizations at the same count level. We first evaluated all the networks on a validation study of which high count reference data was available. We then compared the quantitative performance of all the networks on two patient test studies that included GATE-simulation-inserted liver lesions at different scan durations. The results show that N2NEN training can effectively suppress noise in PET images while preserving natural noise texture, with results similar to the HC target training case while N2N training yielded speckle and clustered noise. N2NEN also outperformed HC target and N2N in terms of lesion quantification.
Paper
Full text
Noise to Noise Ensemble Learning for PET Image Denoising
Semantic Scholar · Medicine · 2019
Abstract
Deep learning convolutional neural networks (DCNN) have been successfully applied to suppress image noise in PET applications. Training a DCNN denoising model usually requires large amounts of paired datasets including high quality targets. In practice, obtaining high quality training targets is often very challenging due to limitations on radiation exposure and imaging times, in addition to the complications of voluntary (e.g. body) and involuntary motions (e.g. respiratory and muscle relaxation). In this study, inspired by the previous work of Noise2Noise training, we hypothesize that instead of training the network to map noisy samples to high quality targets, we could train a similar neural network by pairing one noise realization to an ensemble of noise realizations. We compared a deep residual network trained with three training schemes including one using high count target in training (HC target), one using noise to noise (N2N) training that maps one noise realization to another noise realization at the same count level, and the proposed noise to noise ensemble (N2NEN) training that maps one noise realization to multiple noise realizations at the same count level. We first evaluated all the networks on a validation study of which high count reference data was available. We then compared the quantitative performance of all the networks on two patient test studies that included GATE-simulation-inserted liver lesions at different scan durations. The results show that N2NEN training can effectively suppress noise in PET images while preserving natural noise texture, with results similar to the HC target training case while N2N training yielded speckle and clustered noise. N2NEN also outperformed HC target and N2N in terms of lesion quantification.