Abstract
Variable physiological [(18)F]FDG uptake patterns and a lack of labelled data make it challenging to automatically distinguish normal from pathological suspicious uptake in whole-body PET/CT imaging. We propose a deep learning method that generates patient-specific normal twin PET images to serve as personalized references for quantitative analysis and unsupervised detection of pathological anomalies. We developed an image-to-image generative model that synthesizes normal reference twin PET (ntPET) images from CT scans, patient demographics, and acquisition parameters. The model was trained on 2,538 pseudo-normal PET/CT studies, including stable lymphoma patients and manually disease-masked clinical scans. Model performance was evaluated on 177 test studies achieving 89.3% explained variance and 18.0% mean absolute relative error. We introduced a novel "twin correction" method which reduced SUV(mean) variance by up to 90% in various organs and successfully reduced confounding normally occurring effects of patient sex, age, fat mass, and uptake time. Finally, anomaly detection and unsupervised tumor segmentation was achieved by comparing actual PET scans with their normal twins. The ntPET-based method achieved a dice score of 49.3% on the AutoPET dataset without requiring tumor annotations for training. In conclusion, the proposed ntPET methodology employs personalized normal references to achieve disease-agnostic patient-specific analysis of PET images.