Improving Image Quality and Quantification Accuracy for Static and Dynamic PET

Date of Award

Spring 1-1-2025

Document Type

Dissertation

Degree Name

Doctor of Engineering (DEng)

Department

Biomedical Engineering (ENAS)

First Advisor

Liu, Chi

Abstract

Positron Emission Tomography (PET) is a non-invasive imaging modality widely employed in oncology, neurology, and cardiology to quantitatively assess metabolic and physiological processes. Dynamic PET imaging provides a comprehensive evaluation of tracer kinetics, which is more biologically relevant compared to static imaging alone. In the context of transthyretin cardiac amyloidosis (ATTR-CA), thioflavin-analog tracers such as 18F-flutemetamol have demonstrated diagnostic potential. However, conventional static metrics like the Standardized Uptake Value (SUV) do not fully capture the complex tracer dynamics in the myocardium. This shortfall limits the precise quantification of the amyloid burden and impairs the monitoring of therapy. Additionally, the short frame durations used to capture rapid kinetics yield low-count, noisy images that hinder accurate modeling. While deep learning denoising techniques can improve image quality, challenges remain: supervised methods encounter difficulties in constructing robust training datasets due to inherent variability in injection dose, scan duration, patient physiology, and tracer distribution, which ultimately compromises model generalizability; self-supervised approaches like Deep Image Prior (DIP) avoid this need but rely on static priors that may not capture dynamic changes and require high computational resources for 3D volumes. Moreover, while CycleGAN models enable unpaired image translation for dynamic PET denoising, aligning dynamic frames with a high-quality mean frame can distort critical tracer distributions. Incorporating the full sequence of dynamic frames to preserve kinetic information may mitigate these issues. This thesis introduces advanced kinetic modeling and deep learning denoising techniques to enhance image quality and quantification in static and dynamic PET. In the ATTR-CA study using 18F-flutemetamol, we applied dynamic PET imaging with kinetic modeling to precisely quantify the amyloid burden and assess the effects of six months of tafamidis treatment. Our results showed that the two-tissue reversible compartment models effectively captured tracer kinetics and revealed significant volume of distribution VT reductions correlating with myocardial remodeling and systemic improvements. Second, we addressed data mismatch in supervised deep learning denoising for low-count PET images by examining how training image noise - quantified as the normalized standard deviation in the liver - affects performance, finding that noisier training data improves denoising at the cost of increased spatial blurring. To overcome this trade-off, we developed a personalized method that combines predictions from two networks (one trained on lower-noise inputs and one on higher-noise inputs) using tunable weighting factors, with evaluations on patient and phantom images demonstrating that the optimal weights vary by task. Third, we introduced the Population-based Deep Image Prior (PDIP) for dynamic PET denoising. PDIP used a prior image generated from a supervised model trained on a prompt-matched static PET dataset of 100 clinical studies. We evaluated PDIP on 25%-count and 100%-count dynamic PET images from 23 patients, compared it with a prompt-matched supervised model (PS) and a conditional DIP (CDIP) that used a mean static PET image as prior. While PS and CDIP effectively reduced noise, they tended to smooth images and remove small lesions. CDIP also introduced biased tracer distributions that resulted in lower, inaccurate tracer influx rate (Ki) values. In contrast, PDIP leveraged intrinsic features from both dynamic and large-scale static datasets, achieving comparable noise reduction while enhancing lesion Ki predictions. Last, we proposed the Patlak-Guided Self-Supervised Denoising Network (PG-SSDNet), which is a multi-frame CycleGAN (Cycle-Consistent Generative Adversarial Network) model that integrates a Patlak constraint to preserve tracer kinetics across frames. PG-SSDNet was evaluated on dynamic PET data, consisting of 14-frame and 6-frame sequences from 28 studies, and it outperformed a supervised U-Net, a single-frame CycleGAN, and a version without Patlak guidance by effectively denoising images while maintaining kinetic fidelity. Unlike the DIP model—which, despite better preservation of lesion values, required approximately 4 hours of inference per dynamic study and could not directly process whole-body images due to GPU memory constraints, PG-SSDNet operates without these limitations. In summary, this thesis presented a kinetic modeling method for cardiac amyloidosis and several deep-learning denoising approaches to enhance image quality and quantification accuracy in both static and dynamic PET.

This document is currently not available here.

Share

COinS