Unsupervised Deep Learning with Self-Validation in Dynamic PET Dose Reduction
Lists
Only you see this
Abstract
Dose reduction in dynamic PET imaging confers less radiation exposure however increases noise in reconstructed time frames, which leads to higher noise and possible bias in estimated parametric images. Our aim is to develop self-validated unsupervised deep image prior (DIP) denoising techniques, which do not require vast numbers of well-paired images as the training data, for reduced-dose dynamic PET. We propose two methods with the single-frame (SF-DIP) method processing the dynamic images frame by frame and the all-frame (AF-DIP) method denoising all the frames simultaneously to benefit kinetic modeling. Spatiotemporal information of dynamic imaging is incorporated by using the time-weighted composite image as the DIP input. To benefit from the DIP's capability of capturing image features from corrupted images, the 1/10- dose frames are fed into the network as the target. Setting a hard threshold to optimize the DIP may not work well especially in dynamic PET where the uptake pattern of radiotracers and the noise level of frames change considerably during a scan. Here we adopt a self-validation strategy to ensure the robust performance of DIP. The proposed methods are evaluated using dynamic brain SV2A PET imaging data acquired with <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">11</sup>C-UCB-J. Averaged over thirteen FreeSurfer defined regions, the SF-DIP achieves a noise reduction of 69% relative to the reduced-dose data and a bias of 0.1% with respect to the full-count data. The noise reduction and bias for the AF-DIP method are 72% and -1.6%, respectively. Such a noise reduction (bias) translates into 51% (-2.3%) in BP<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ND</inf> images after kinetic modeling for the SF-DIP and 57% (-0.6%) for the AF-DIP. To conclude, the proposed unsupervised-learning based denoising methods demonstrate better performance in noise reduction while keeping bias smaller compared with other denoising methods. Processing all time frames simultaneously further improves the performance of the unsupervised method in kinetic modeling analysis.
BibTeX
@misc{li2023unsupervised,
author = {Adam Li and M.R. Syed and Mika Naganawa and D. Matuskery and Richard E. Carson and Jing Tang},
title = {{Unsupervised Deep Learning with Self-Validation in Dynamic PET Dose Reduction}},
year = {2023},
doi = {10.1109/nssmicrtsd49126.2023.10337886},
}Loading related content...