Populational and individual information based PET image denoising using conditional unsupervised learning

Abstract

Our study aims to improve the signal-to-noise ratio of positron emission tomography (PET) imaging using conditional unsupervised learning. The proposed method does not require low- and high-quality pairs for network training which can be easily applied to existing PET/computed tomography (CT) and PET/magnetic resonance (MR) datasets. This method consists of two steps: populational training and individual fine-tuning. As for populational training, a network was first pre-trained by a group of patients’ noisy PET images and the corresponding anatomical prior images from CT or MR. As for individual fine-tuning, a new network with initial parameters inherited from the pre-trained network was fine-tuned by the test patient's noisy PET image and the corresponding anatomical prior image. Only the last few layers were fine-tuned to take advantage of the populational information and the pre-training efforts. Both networks shared the same structure and took the CT or MR images as the network input so that the network output was conditioned on the patient’s anatomic prior information. The noisy PET images were used as the training and fine-tuning labels. The proposed method was evaluated on a 68Ga-PPRGD2 PET/CT dataset and a 18F-FDG PET/MR dataset. For the PET/CT dataset, with the original noisy PET image as the baseline, the proposed method has a significantly higher contrast-to noise ratio (CNR) improvement (71.85% ± 27.05%) than Gaussian (12.66% ± 6.19%, P = 0.002), nonlocal mean method (22.60% ± 13.11%, P = 0.002) and conditional deep image prior method (52.94% ± 21.79%, P = 0.0039). For the PET/MR dataset, compared to Gaussian (18.73% ± 9.98%, P < 0.0001), NLM (26.01% ± 19.40%, P < 0.0001) and CDIP (47.48% ± 25.36%, P < 0.0001), the CNR improvement ratio of the proposed method (58.07% ± 28.45%) is the highest. In addition, the denoised images using both datasets also showed that the proposed method can accurately restore tumor structures while also smoothing out the noise.

Positron Range Corrections and Denoising Techniques for Gallium-68 PET Imaging: A Literature Review

“…Two different approaches were identified for reducing the noise in 68 Ga PET images and enabling low-count 68 Ga PET measurements. The first approach reduces the noise during the image reconstruction process (reconstruction-based noise reduction approaches, n = 11) [ 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 ], whereas the second approach is based on neural networks (deep learning approaches for noise reduction, n = 5) [ 37 , 38 , 39 , 40 , 41 ], as seen in Figure 2 .…”

Unsupervised arterial spin labeling image superresolution via multiscale generative adversarial network

“…To solve this problem, fine-tuning based on a model pretrained by population data can help, similar to what we have done for PET denoising. 57 Based on a population-trained model, fine-tuning only takes 1 min. Further improving the ASL SR model based on both individual and populational datasets to improve its robustness and reduce the training time is one of our future works.…”

“…Our current implementation takes around 3 h for LR‐to‐NR training and around 7 h for NR‐to‐SR training, which is much longer than the scan time of high‐resolution pCASL (44 min). To solve this problem, fine‐tuning based on a model pretrained by population data can help, similar to what we have done for PET denoising 57 . Based on a population‐trained model, fine‐tuning only takes 1 min.…”

Deep learning-based image reconstruction and post-processing methods in positron emission tomography for low-dose imaging and resolution enhancement

“…DIP-based methods produced better contrast-to-noise ratio compared to the deep decoder method from Heckel et al [ 105 ], Gaussian smoothing, non-local means and 4D block matching. Work from Cui et al [ 106 ] extended this concept by implementing the DIP method with a neural network initially trained on population level information, demonstrating better results than a randomly initialised DIP. Furthermore, Yie et al [ 78 ] investigated the quality of supervised learning approaches with noisy target images.…”