DosimeTron: Automating Personalized Monte Carlo Radiation Dosimetry in PET/CT with Agentic AI
Summary: arXiv:2604.06280v1 Announce Type: cross
Abstract
The field of medical imaging is witnessing a transformative shift with the introduction of advanced AI technologies. One such innovation is DosimeTron, an agentic AI system developed for the automation of patient-specific Monte Carlo (MC) internal radiation dosimetry in PET/CT examinations. This article presents a comprehensive overview of the system’s purpose, methodology, results, and implications for clinical practice.
Purpose
The primary objective of DosimeTron is to enhance the accuracy and efficiency of internal radiation dosimetry for patients undergoing PET/CT scans. By leveraging advanced AI capabilities, the system aims to streamline the dosimetry process, thereby improving patient outcomes and optimizing treatment plans.
Materials and Methods
In a retrospective study, DosimeTron was rigorously evaluated using a publicly available PSMA-PET/CT dataset, which included 597 studies from 378 male patients. These studies were acquired on three different scanner models, comprising 369 instances of 18-F and 228 of 68-Ga. The system employs GPT-5.2 as its reasoning engine and integrates 23 tools through four Model Context Protocol servers. This architecture allows for:
- DICOM metadata extraction
- Image preprocessing
- Monte Carlo simulation
- Organ segmentation
- Dosimetric reporting via natural-language interaction
The agentic performance of DosimeTron was assessed using various prompt templates, ranging from single-turn instructions to multi-turn conversational exchanges. Monitoring was conducted via OpenTelemetry traces to ensure system reliability.
Results
The evaluation yielded impressive results, with no execution failures, pipeline errors, or hallucinated outputs across all prompt templates and runs. Key findings included:
- Pearson’s correlation coefficient (r) was observed to range from 0.965 to 1.000, with a median of 0.997 and all p-values < 0.001.
- Lin’s concordance correlation coefficient (CCC) ranged from 0.963 to 1.000, with a median of 0.996.
- The mean absolute percentage difference was below 5% for 19 out of 22 organs, with a median difference of 2.5%.
- The total per-study processing time averaged 32.3 minutes with a standard deviation of 6.0 minutes.
Conclusion
DosimeTron has demonstrated its capability to autonomously execute complex dosimetry pipelines across various prompt configurations. The high dosimetric agreement achieved with OpenDose3D, coupled with clinically acceptable processing times, underscores the potential of agentic AI in revolutionizing patient-specific Monte Carlo dosimetry in PET/CT. This development not only enhances the accuracy of radiation dose calculations but also paves the way for more personalized and effective treatment strategies in nuclear medicine.