A novel methodology to train and deploy a machine learning model for personalized dose assessment in head CT.

Abstract

To propose a machine learning-based methodology for the creation of radiation dose maps and the prediction of patient-specific organ/tissue doses associated with head CT examinations. CT data were collected retrospectively for 343 patients who underwent standard head CT examinations. Patient-specific Monte Carlo (MC) simulations were performed to determine the radiation dose distribution to patients' organs/tissues. The collected CT images and the MC-produced dose maps were processed and used for the training of the deep neural network (DNN) model. For the training and validation processes, data from 231 and 112 head CT examinations, respectively, were used. Furthermore, a software tool was developed to produce dose maps from head CT images using the trained DNN model and to automatically calculate the dose to the brain and cranial bones. The mean (range) percentage differences between the doses predicted from the DNN model and those provided by MC simulations for the brain, eye lenses, and cranial bones were 4.5% (0-17.7%), 5.7% (0.2-19.0%), and 5.2% (0.1-18.9%), respectively. The graphical user interface of the software offers a user-friendly way for radiation dose/risk assessment. The implementation of the DNN allowed for a 97% reduction in the computational time needed for the dose estimations. A novel methodology that allows users to develop a DNN model for patient-specific CT dose prediction was developed and implemented. The approach demonstrated herein allows accurate and fast radiation dose estimation for the brain, eye lenses, and cranial bones of patients who undergo head CT examinations and can be used in everyday clinical practice. • The methodology presented herein allows fast and accurate radiation dose estimation for the brain, eye lenses, and cranial bones of patients who undergo head CT examinations and can be implemented in everyday clinical practice. • The scripts developed in the current study will allow users to train models for the acquisition protocols of their CT scanners, generate dose maps, estimate the doses to the brain and cranial bones, and estimate the lifetime attributable risk of radiation-induced brain cancer.