ABSTRACT
Radiation-induced toxicity can be categorized according to its onset time into early and late effects. Toxicity of tissue may relate to dose and fractionation, tissue structure and architecture and deoxyribonucleic acid repair ability. The former approach mathematically formulates toxicity based on simplified biophysical understanding of radiation effects on cells primarily from in vitro cell culture experiments. In the context of machine learning, the prediction of normal tissue complication probability (NTCP) can be viewed as a supervised learning problem. Although, NTCP modeling is largely a supervised learning problem, some unsupervised learning approaches such as principal component analysis can be applied before supervised learning to improve performance and increase robustness. A pure accuracy criterion is simple, but can cause problems when the dataset is imbalanced, a common situation in NTCP prediction. For evaluation and validation of NTCP models, realization of testing the model on some external sources can be an issue.
