Revolutionizing Internal Dosimetry with AI: PET-Only Solutions for Precision Medicine

Abstract

The application of artificial intelligence (AI) in positron emission tomography (PET)-based internal dosimetry in radiopharmaceutical therapy (RPT), with an emphasis on the paradigm shift from resource-consuming multi-time-point acquisition protocols toward efficient data-driven dosimetric frameworks. Machine learning (ML) models capable of determining organ- and patient-specific absorbed dose estimates from minimal imaging data are the focus of this review. The chapter illustrates how pre-therapy PET features, combined with readily available clinical variables, can aid in effective half-life prediction and early post-therapy dose estimation, as seen in recent instant single-time-point (iSTP) dosimetry frameworks. Deep learning architectures like U-Net variants, convolutional neural networks, and transformer-based models are tested to mimic Monte Carlo-level dose accuracy at a fraction of the computational cost. AI contributions reviewed include sparse-view and low-count SPECT reconstruction, automated multi-organ segmentation, cross-modality image synthesis, and pre-therapy predictive dosimetry using theranostic imaging pairs. We discuss methodologies, multi-center validation, regulatory issues, and implications for precision medicine. The evidence collected in this review indicates that AI-based PET dosimetry is a clinically viable pathway toward personalized radiopharmaceutical therapy that is quantitatively accurate and operationally feasible within routine clinical workflows.