Study On Patient-specific Radiation Dosimetry Method Based On PET/CT Multi-organ Automatic Segmentation And Fast Monte Carlo Calculation

Radiological imaging techniques represented by CT and PET/CT have been widely used in medical diagnosis.However,the extensive use of radiological medical imaging examinations quickly raised concerns about the risk of radiation-induced cancer in multiple imaging.For these risks,it is necessary to calculate the absorbed dose to the patient’s organs.The existing dose calculation tools can only be used to evaluate the radiation level received by the general population,and are not sufficient for individualized organ dose quantification in clinical practice.The main challenges are reflected in two aspects:One is the rapid construction of personalized phantom,and the other is the rapid and accurate calculation of radiation doses.Researchers need to develop personalized organ dose calculation tools with sufficient speed and accuracy to be used in routine clinical practice.Aiming at the above problems,this paper aims to propose and verify a multi-organ automatic segmentation and fast Monte Carlo personalized radiation dose calculation method based on PET/CT image data.The specific research goal is to use CT images to construct an automatic organs segmentation model based on neural network to complete the rapid design of personalized phantoms,and to develop GPU-accelerated Monte Carlo radiation dose calculation software to achieve rapid radiation dose calculation in radiological medical imaging.The main work and research results of this paper are as follows:(1)Research on the rapid construction of personalized phantoms based on automatic segmentation of multiple organs.Automatic organ segmentation is a good method for building personalized phantoms.However,it is extremely difficult to train a segmentation model for whole-body organ.This paper proposes a set of automatic segmentation methods for whole body organs that are classified first and then segmented.Specifically,a CT classification network model is trained based on the EfficientNet architecture,and the whole body is divided into 5 categories(head and neck,chest,abdomen,pelvis and legs).On the basis of 3D U-Net,a residual module and a deep supervision module are added,and the downsampling layer is adjusted according to the size of the input CT image sample to build 3D ResU-Net.Then,multiorgan segmentation network models such as head and neck,chest,abdomen,and pelvis,and single-organ segmentation models such as brain,lens,skin,and spinal cord were trained respectively.For the organ segmentation of a new patient,the CT images are first divided into different parts categories,the CT images of each part are input into the corresponding organ segmentation network for prediction,and finally the segmentation results of all organs are merged.In 50 test patients(over 4000 CT images),the overall classification accuracy of the CT classification network model exceeded 99%,the Dice coefficient of the organ segmentation network model for most organs in the body was at or close to 0.9,and personalized phantom construction of a patient with whole body PET/CT imaging takes about 1 minute.(2)Monte Carlo radiation dose calculation based on GPU acceleration.Combining the personalized phantom construction method and a GPU-accelerated Monte Carlo calculation engine ARCHER,the personalized radiation dose calculation software ARCHER-CT for CT imaging was developed,and its dose results were compared with those of VirtualDose,a commercial software for CT dose based on a reference human model.In addition,by building a radioactive source module related to nuclear medicine(NM)radionuclides,a Monte Carlo software ARCHER-NM for individualized radiation dose calculation for PET imaging was developed,and it dose results were compared with those of GATE,a popular CPU-based Monte Carlo software.The results show that the ARCHER-CT dose calculation error is more than 10 times lower than that of the VirtualDose,and the ARCHER-NM is more than 800 times faster than GATE,and the average relative error of the organ dose calculated by the two is not more than 2%.(3)Clinical application of PET/CT personalized radiation dose calculation software.This paper presents a method for calculating the time-cumulative activity distribution of PET/CT imaging patients so as to calculate the internal radiation dose of PET imaging,and a personalized phantom program interface and.a PET/CT imaging personalized radiation dose calculation program interface are developed to embed them into the clinical radiation dose calculation software DeepPlan,and a clinical test of individualized radiation dose calculation in 10 PET/CT patients was carried out.The CT imaging dose calculation takes an average of only 48 seconds per patient,and the PET imaging dose calculation takes an average of 22 seconds per patient.The research results show that the multi-organ automatic segmentation and fast Monte Carlo personalized radiation dose calculation method based on PET/CT image data proposed in this paper is feasible,and the PET/CT imaging personalized radiation dose calculation software developed in this study has sufficient speed and accuracy,and provides an effective tool for personalized radiation dose assessment of clinical PET/CT imaging patients,and also provides an effective solution for personalized radiation dose calculation in other radiological medical imaging applications.