Title:A Method of Rapid Quantification of Patient-Specific Organ Dose for CT Using Coupled Deep-Learning based Multi-Organ Segmentation and GPU-accelerated Monte Carlo Dose Computing

Abstract:Purpose: This paper describes a new method to apply deep-learning algorithms for automatic segmentation of radiosensitive organs from 3D tomographic CT images before computing organ doses using a GPU-based Monte Carlo code. Methods: A deep convolutional neural network (CNN) for organ segmentation is trained to automatically delineate radiosensitive organs from CT. With a GPU-based Monte Carlo dose engine (ARCHER) to derive CT dose of a phantom made from a subject's CT scan, we are then able to compute the patient-specific CT dose for each of the segmented organs. The developed tool is validated by using Relative Dose Error (RDE) against the organ doses calculated by ARCHER with manual segmentation performed by radiologists. The dose computation results are also compared against organ doses from population-average phantoms to demonstrate the improvement achieved by using the developed tool. In this study, two datasets were used: The Lung CT Segmentation Challenge 2017 (LCTSC) dataset, which contains 60 thoracic CT scan patients each with 5 segmented organs, and the Pancreas-CT (PCT) dataset, which contains 43 abdominal CT scan patients each with 8 segmented organs. Five-fold cross-validation of the new method is performed on both datasets. Results: Comparing with the traditional organ dose evaluation method that based on population-average phantom, our proposed method achieved the smaller RDE range on all organs with -4.3%~1.5% vs -31.5%~33.9% (lung), -7.0%~2.3% vs -15.2%~125.1% (heart), -18.8%~40.2% vs -10.3%~124.1% (esophagus) in the LCTSC dataset and -5.6%~1.6% vs -20.3%~57.4% (spleen), -4.5%~4.6% vs -19.5%~61.0% (pancreas), -2.3%~4.4% vs -37.8%~75.8% (left kidney), -14.9%~5.4% vs -39.9% ~14.6% (gall bladder), -0.9%~1.6% vs -30.1%~72.5% (liver), and -23.0%~11.1% vs -52.5%~-1.3% (stomach) in the PCT dataset.