Novel Insights for Radiation Risk Assessment Unveiled by Deep Learning

Abstract

AbstractContemporary radiation risk assessment predominantly depends on nonlinear parametric models, which typically include a baseline term, a dose-response term, and an effect modifier term. Despite their widespread application in estimating tumor risks, parametric models face a notable drawback: their rigid model structure can be overly restrictive, potentially introducing bias and inaccuracies into risk estimations.In this study, we analyze data on solid tumors and leukemia from the Life Span Study (LSS) to compare the performance of deep neural network (DNN) and nonlinear parametric (NLP) models in assessing ERRs. DNN presents novel perspectives for radiation risk assessment. Our findings indicate that DNN can perform better than the traditional parametric models. Even if DNN and NLP models exhibit similar performance in predicting tumor incidence, they diverge significantly in their estimated ERRs. Standard NLP models tend to underestimate ERRs directly linked to radiation dose, overestimate ERRs for individuals at younger attained ages and ages at exposure, and underestimate ERRs for those at older attained ages. Furthermore, DNN consistently identifies radiation dose as the primary and predominant risk factor for ERRs in leukemia and solid tumors, underscoring the critical role of radiation dose in risk assessment. The insights from DNN could enhance low-dose radiation risk assessment and improve parametric model development.