Systematic investigation of artifacts in deep learning-based PET image enhancement

About This Grant

Project Summary/Abstract Deep learning (DL)-based image denoisers have gained significant attention in positron emission tomography (PET) due to their ability to enhance image quality in low-count or short-duration scans. These models hold substantial promise for reducing radiation dose, minimizing scan time, and improving diagnostic performance – particularly for vulnerable populations and in resource-limited settings. However, DL-based denoisers may introduce artifactual features, including false-positive or false-negative lesions, that are not supported by the underlying data. These artifacts remain poorly understood and pose a critical barrier to the safe clinical adoption of AI-enhanced image reconstruction. This project aims to systematically evaluate, quantify, and model the risk of artifacts in DL-denoised PET images. Using a Monte Carlo-based lesion embedding framework developed by our team (DIANA), we will simulate artificial hepatic lesions in real human PET data to establish in-vivo ground truth. Low-count images of various noise levels will be generated through random downsampling of high-quality PET data acquired on the EXPLORER total-body PET/CT scanner. These images will subsequently be denoised using a state-of-the-art 3D diffusion probabilistic model, and artifacts will be detected using a novel combination of gradient-domain image differencing and generalized scan statistics. Furthermore, we will systematically investigate how imaging and anatomical factors that are available at scan time influence the probability of wrongfully added lesions in DL-denoised PET images. To that end we will develop a logistic regression model to quantify the conditions under which artifacts are most likely to occur, thus, providing a prototype predictive framework for risk stratification and clinical decision support. Input features will include image noise level as well as anatomical and patient-specific variables (e.g., BMI, and amount of injected radioactivity). This work enables a scientific assessment of the safety and fidelity of AI-based denoising methods, ensuring they support rather than compromise clinical decision-making and the broader goals of translational science.