Patient-specific Monte Carlo Dose Reconstruction in Whole-body CT Using Deep Neural Networks

Overview

This study developed a fully automated deep learning method to estimate patient-specific Monte Carlo (MC) dose maps for whole-body CT scans in real time without requiring real-time acquisition parameters. The approach demonstrated accurate dose reconstruction using patient CT images and density maps, addressing limitations of conventional MC simulations and existing software tools.

Background

Computed tomography (CT) is widely used for diagnosis, follow-up, and therapy planning but contributes significantly to patient ionizing radiation exposure. Accurate estimation of organ doses is critical for personalized radiation risk assessment and optimization following ALARA principles. Monte Carlo simulations provide the most accurate dose calculations but are computationally intensive and complex, limiting clinical adoption. Deep learning offers potential to accelerate and automate dose estimation using patient-specific imaging data.

Data Highlights

The study included 63 patients undergoing whole-body PET/CT scans with 120 kVp tube voltage and Siemens CareDose4D tube current modulation. Monte Carlo simulations were performed for each patient to generate 3D dose maps, modeling helical scanning with multiple discrete source positions. Additional simulations at 90 kVp were conducted for testing network generalizability. A uniform water-filled cylinder phantom was also simulated at both 90 and 120 kVp to validate dose calculations.

Key Findings

• Deep neural networks were trained to predict patient-specific MC-based dose maps from CT images and density maps without requiring real-time acquisition parameters.
• MC simulations incorporated detailed modeling of scanning parameters including tube voltage, collimation, table speed, rotation time, pitch, and tube current modulation.
• The method achieved real-time dose estimation suitable for clinical workflow integration, overcoming the computational burden of traditional MC simulations.
• Testing with 90 kVp data and uniform phantom simulations demonstrated the model's generalizability and accuracy across different tube voltages.
• Automated body contour segmentation and density mapping facilitated streamlined input preparation for dose reconstruction.

Clinical Implications

This deep learning-based approach enables rapid, patient-specific radiation dose estimation during whole-body CT imaging without the need for detailed acquisition parameters at scan time. It supports personalized radiation risk assessment and optimization, potentially improving adherence to ALARA principles in clinical practice. The method may facilitate routine organ dose monitoring and dose management in serial CT examinations.

Conclusion

The study successfully developed and validated a deep neural network framework for patient-specific Monte Carlo dose reconstruction in whole-body CT imaging, providing accurate, real-time dose maps without requiring real-time acquisition parameters. This advancement holds promise for enhancing personalized radiation dose assessment and optimization in clinical settings.

References

1. Schneider et al. 2000 -- Conversion of CT Hounsfield units to tissue density
2. International Commission on Radiological Protection (ICRP) -- Radiation dose estimation and ALARA principles
3. Siemens Healthineers Biograph mCT -- Whole-body PET/CT imaging
4. MCNPX Monte Carlo code version 2.6 -- Radiation transport simulation