Affiliations: [a]
Institute of Radiation Medicine, Fudan University, Shanghai, China
| [b]
Department of Nuclear Science and Technology, Institute of Modern Physics, Fudan University, Shanghai, China
| [c] Key Lab of Nuclear Physics & Ion-Beam Appl. (MOE), Fudan University, Shanghai, China
| [d]
Department of Radiation Oncology, Shanghai Jiao Tong University Chest Hospital Shanghai, China
| [e]
Department of Radiology, Shanghai Zhongye Hospital, Shanghai, China
Correspondence:
[*]
Corresponding author: Haikuan Liu, Institute of Radiation Medicine, Fudan University, Shanghai, China. E-mail: [email protected].
Note: [1] Co-first authors: The authors contributed equally to this work.
Note: [2] Co-corresponding authors: These authors contributed equally as co-corresponding authors.
Abstract: PURPOSE:
This study aims to propose and develop a fast, accurate, and robust prediction method of patient-specific organ doses from CT examinations using minimized computational resources. MATERIALS AND METHODS:
We randomly selected the image data of 723 patients who underwent thoracic CT examinations. We performed auto-segmentation based on the selected data to generate the regions of interest (ROIs) of thoracic organs using the DeepViewer software. For each patient, radiomics features of the thoracic ROIs were extracted via the Pyradiomics package. The support vector regression (SVR) model was trained based on the radiomics features and reference organ dose obtained by Monte Carlo (MC) simulation. The root mean squared error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R-squared) were evaluated. The robustness was verified by randomly assigning patients to the train and test sets of data and comparing regression metrics of different patient assignments. RESULTS:
For the right lung, left lung, lungs, esophagus, heart, and trachea, results showed that the trained SVR model achieved the RMSEs of 2 mGy to 2.8 mGy on the test sets, 1.5 mGy to 2.5 mGy on the train sets. The calculated MAPE ranged from 0.1 to 0.18 on the test sets, and 0.08 to 0.15 on the train sets. The calculated R-squared was 0.75 to 0.89 on test sets. CONCLUSIONS:
By combined utilization of the SVR algorithm and thoracic radiomics features, patient-specific thoracic organ doses could be predicted accurately, fast, and robustly in one second even using one single CPU core.
