ALAndi Li

Andi Li

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Andi Li has a total of 23 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    NA-DIP: A Noise-Aware Deep Image Prior for Generating Dedicated Brain PET–Quality Images from Whole-Body PET Acquisitions

    by Chenyang Han, Andrew T. Trout, Andi Li, Joseph G. Meier, Cara E. Morin, Adam F. Prasanphanich, Susan E. Sharp, Christopher G. Anton, Jing Tang

  • Open Access

    Acquisition time/dose reduction in pediatric PET imaging using patch-based deep learning

    by Chenyang Han, Andrew T. Trout, Andi Li, Joseph G. Meier, Vinicius P. V. Alves, Joseph MacLean, Nadeen Abu Ata, Yinan Li, Susan E. Sharp, Christopher G. Anton, et al.

    Deep learning (DL)-based denoising methods have shown promise for reducing radiation dose and/or acquisition time in pediatric PET imaging. However, conventional DL approaches typically require large and diverse training datasets to achieve generalizability. We propose a patch-based DL (PDL) approach that learns local structural representations from a single high-quality examination, with the goal of improving image quality and preserving quantitative accuracy of reduced-count pediatric whole-body PET while minimizing training burden. This retrospective study included PET/CT datasets acquired on a GE Discovery MI Gen 2 scanner. A single pediatric examination with high structural conspicuity and low noise was used for training. Testing was performed using a NEMA phantom and 88 clinical pediatric examinations (age < 12 years) acquired with routine pediatric FDG dosing and a 90‑s‑per‑bed protocol. Clinical list‑mode data were truncated to 20 s per bed to simulate reduced‑count acquisitions; training and phantom datasets were downsampled to match this noise level. The PDL model was trained using paired full‑count and reduced‑count image patches and applied to denoise reduced‑count phantom and patient images. Phantom evaluations assessed spatial resolution, contrast recovery, and noise. Clinical evaluation included a blinded observer study and quantitative SUV analysis, followed by assessment of the noise-bias tradeoff. PDL preserved spatial resolution and contrast while achieving substantial noise reduction in phantom experiments. In the observer study, PDL‑enhanced 20‑s‑per‑bed images were rated as equivalent or superior in overall image quality and noise relative to standard 90‑s‑per‑bed images in more than 85% of cases. Small negative relative bias in SUVmax and SUVmean was observed in the liver, mediastinal blood pool, and lesions, with no statistically significant differences compared with standard‑count images. Compared with highly regularized reconstruction, PDL achieved greater noise reduction with reduced bias. PDL enhancement for pediatric whole-body PET imaging may enable a reduction in acquisition time to 20 s per bed or a reduction in administered activity to approximately 22% of current clinical practice, while maintaining spatial resolution, contrast, quantitative accuracy, and diagnostic image quality.


  • Dose Reduction in Dynamic PET Imaging Using Self-Supervised Two-Step Deep Image Prior

    by Andi Li, Mohammad B Syed, Mika Naganawa, David Matuskey, Richard E Carson, Jing Tang

    Reducing the radiotracer dose in dynamic brain PET imaging lowers patient radiation exposure but also degrades the signal-to-noise (SNR) ratio and introduces bias in both dynamic frames and resulting parametric images. These effects compromise the accuracy of quantitative assessments of target proteins. In this study, we propose a self-supervised two-step deep image prior (TS-DIP) method that integrates composite image-based pretraining with individual frame-based fine-tuning to enhance reduced-dose dynamic brain PET imaging. We evaluated the TS-DIP method on 1/10-count dynamic datasets acquired using two tracers with distinct uptake patterns, 11C-UCB-J (targeting synaptic density) and 11C-LSN3172176 (targeting muscarinic cholinergic receptors). Compared with conventional, supervised deep learning, and conditional DIP denoising approaches, the TS-DIP demonstrated a superior tradeoff between noise reduction and bias in both dynamic frames and parametric images. Averaged across all dynamic frames of 11C-UCB-J subjects, the proposed method achieved a 75% reduction in noise with an bias of 2.1%, corresponding to 63% and 60% noise reduction in the K1 (bias, -1.2%) and VT (bias, -3.1%) parametric images, respectively. For 11C-LSN3172176 subjects, the average noise reductions (bias) were 80% (3.4%) in dynamic frames, 62%(-0.38%) in K1 images, and 77% (-3.4%) in VT images. These results highlight the generalizability of the proposed TS-DIP method across tracers with varying activity distributions and its potential to enable accurate, reduced-dose dynamic brain PET imaging without the need for large training datasets.


  • Open Access

    Dose reduction for synaptic density PET imaging in Parkinson’s disease

    by Andi Li, Mika Naganawa, Praveen Honhar, Kathryn Fontaine, Paul Gravel, David Matuskey, Richard E. Carson, Jing Tang


  • Deep Image Prior-Incorporated Direct PET Parametric Image Reconstruction

    by Andi Li, Mohammad B. Syed, Jonathan B. Moody, Jing Tang

    Direct parametric reconstruction algorithms have been developed to improve the statistical reliability of parametric images estimated from dynamic PET imaging data. However, these estimates are degraded by noise due to measurement error and noise propagation during reconstruction. In this study, we develop a deep image prior (DIP) regularized direct reconstruction method, where the DIP network is used to represent the estimated parametric image. By initializing the DIP with pre-trained weights and updating its network to learn the intermediate information during reconstruction, the DIP regularization leverages the available population and subject-specific features. The proposed method is applied to reconstruct K1 from both simulated and patient data acquired by 82Rb dynamic PET myocardial perfusion imaging. Benefiting from the nonlinear representation capability of the DIP network, the proposed method achieves superior noise versus bias/mean performance compared with the indirect and direct reconstruction methods with various regularizations formed by quadratic smoothness, dictionary learning, or fully-connected neural network. To summarize, the proposed method demonstrates its potential in improving the precision of dynamic PET imaging measurements, which will contribute to diagnostic accuracy and disease monitoring.