Using Medical Imaging Effective Dose in Deep Learning Models: Estimation and Evaluation

Accurately estimating patient exposure is a fundamental concern when modeling the radiation risk from medical imaging. Inaccurate estimation techniques produce misleading instances that have a cascading effect on the performance of risk models. A commonly used method of estimating exposure is using mean effective dose (ED) values from the literature. However, the predictive power of literature values to impute patients ED has not been investigated. In this article, we present a comparative analysis between two methods to estimate ED for computed tomography (CT) and X-ray (XR) scans: 1) mean EDs from the literature and 2) calculated dose estimates from imaging scans using ED tools. We also used adversarial machine learning to demonstrate how the difference between estimation methods impacts a proof-of-concept deep learning model. The study cohort had 39 909 medical imaging scans (7427 CT and 32 482 XR scans) from a stratified random sample of 2000 patients from four hospitals in Hamilton, Canada over ten years. Our results showed moderate increases in the mean ED compared to the literature across all exam types. However, using mean values reported in the literature underestimated patients' total ED over the study period. Our results also demonstrated that the differences between the estimation methods were enough to cause model misclassifications. The results of our study demonstrate the challenges of using mean ED values from the literature to estimate patient medical imaging exposure. There is a need to develop novel imputation methods to estimate patients' EDs from medical imaging.