A method of rapid quantification of patient‐specific organ doses for CT using deep‐learning‐based multi‐organ segmentation and GPU‐accelerated Monte Carlo dose computing

Zhao Peng, Xi Fang, Pingkun Yan, Hongming Shan, Tianyu Liu, Xi Pei, Ge Wang, Bob Liu, Mannudeep K. Kalra, X. George XuMedical 2020 · DOI · arXiv

Abstract

This work shows the feasibility to perform combined automatic patient-specific multi-organ segmentation of CT images and rapid GPU-based Monte Carlo dose quantification with clinically acceptable accuracy and efficiency.

Introduction

Computed Tomography (CT) scans are indispensable in modern medicine, yet they expose patients to ionizing radiation, which carries a cumulative risk of cancer. Managing this risk requires precise knowledge of the "organ dose"—the energy absorbed by specific radiosensitive organs such as the lungs, heart, or liver.

Historically, clinical dose reporting has relied on "population-average" phantoms—standardized mathematical models of the human body. However, because these models do not account for individual anatomical variations, dose estimates can be inaccurate by over 100%. While patient-specific dosimetry is the gold standard for accuracy, it has traditionally been hindered by the time-consuming nature of manual organ segmentation and computationally intensive Monte Carlo simulations.

Methods

To overcome the speed-accuracy trade-off, the authors proposed a parallel computational workflow that integrates deep learning with high-performance computing. The model was trained and validated using the LCTSC (Lung CT Segmentation Challenge) and PCT (Pancreas-CT) datasets, covering thoracic and abdominal regions.

Overview of the proposed parallel workflow for organ segmentation and dose calculation.

Automatic Segmentation

The workflow utilizes a 3D U-Net convolutional neural network (CNN) to automatically delineate organs from raw CT images. To improve the segmentation of small or narrow structures like the esophagus and spinal cord, a hierarchical "fine segmentation" step was implemented, where these regions are cropped and re-processed at higher resolutions.

Architecture of the 3D U-Net used for multi-organ segmentation.

Rapid Dose Calculation

Simultaneously, the ARCHER GPU-accelerated Monte Carlo code calculates the radiation dose. ARCHER simulates the transport of millions of X-ray photons through the patient’s specific anatomy, accounting for physical interactions like the photoelectric effect and Compton scattering. By running segmentation and dose simulation in parallel, the total processing time is significantly reduced.

Results and Key Insights

The proposed method achieved a total processing time of less than 10 seconds, transforming dosimetry from a retrospective report into a potential prospective tool for scan optimization.

  • Segmentation Accuracy: The network achieved high Dice Similarity Coefficients (DSC), with medians of 0.97 for the right lung and 0.96 for the liver. Performance was lower for the duodenum (0.64) due to poor soft-tissue contrast.
  • Dose Precision: The Relative Dose Error (RDE) was significantly lower than phantom-based methods. For instance, the error range for heart dose was narrowed from -15.2% to 125.1% (using standard phantoms) to just -7.0% to 2.3% using the patient-specific approach.
  • Efficiency: Automatic segmentation takes less than 5 seconds, and ARCHER calculates the dose for 100 million photons in less than 4 seconds on an Nvidia Titan RTX GPU, maintaining a statistical uncertainty of less than 0.5%.

Discussion and Future Work

While this research marks a significant step toward personalized radiation safety, some challenges remain. The performance on smaller organs was limited by training data scarcity. Future work will focus on expanding the model to include more diverse datasets, handling anatomical abnormalities (e.g., tumors or implants), and extending the scope to pediatric cases and other body regions like the head and neck.

Ultimately, this work demonstrates that the computational barriers to real-time, patient-specific radiation dosimetry are rapidly diminishing.

Summary updated

Abstract

This work shows the feasibility to perform combined automatic patient-specific multi-organ segmentation of CT images and rapid GPU-based Monte Carlo dose quantification with clinically acceptable accuracy and efficiency.
BibTeX
@inproceedings{peng2020method,
  author = {Zhao Peng and Xi Fang and Pingkun Yan and Hongming Shan and Tianyu Liu and Xi Pei and Ge Wang and Bob Liu and Mannudeep K. Kalra and X. George Xu},
  title = {{A method of rapid quantification of patient‐specific organ doses for CT using deep‐learning‐based multi‐organ segmentation and GPU‐accelerated Monte Carlo dose computing}},
  booktitle = {Medical},
  year = {2020},
  doi = {10.1002/mp.14131},
}
Figure