Auto-PET-IQA: A Fully Automated Region-Specific PET Image Quality Assessment Tool

Incorporating fully automated human perception-driven PET image quality assessment (IQA) measures into the scanning routine is beneficial due to: (1) it could be useful in a clinical setting to instantly identify the need for reperformed acquisition or image reconstruction; (2) by delivering high-quality scans, it alleviates radiologists’ daily workload and expedites the diagnosis and monitoring procedure; and (3) low image quality is a primary limitation in the development of different artificial intelligence (AI)-based models for PET processing and analysis, and an automated PET IQA model can serve as a useful tool to organize and refine input data towards AI-based models for PET analysis. In this study, we developed fully automated AI-based region-specific [¹⁸F]FDG-PET IQA models. The IQ of 174 ¹⁸F-PET/CT scans (acquired on two scanners) were rated by two nuclear medicine physicians and binarized to low- and high-quality. Region identification (head and neck, chest, chest-abdomen interval, abdomen, and pelvis) and body contouring (BC) (excluding the arms) were automatically performed utilizing deep learning (DL) models on CT images and then edited/verified by the physicians. For radiomics-machine learning (R-ML) models, IBSI-based features were fed into multiple feature selection and ML algorithms. DL model was developed using the SEResNet50 architecture. All R-ML and DL models were evaluated using AUC, accuracy, sensitivity, and specificity. Most effected features selected in radiomics models are reported. Except for the chest-abdomen region, DL model performance was superior to R-ML models. Region-wise AUCs of DL vs. best R-ML model were: head - 0.91 vs. 0.85, chest - 0.82 vs. 0.79, chest-abdomen - 0.7 vs. 0.77, abdomen - 0.8 vs. 0.79, and pelvis - 0.91 vs. 0.75, respectively.