ABSTRACT
Most of the normal tissue complication probability models are based on the dose-volume histograms (DVH). Dimensionality reduction, feature extraction strategies together with machine learning methodologies aimed at exploiting more available multimodal data have emerged to overcome some of these issues, exhibiting promising prediction capabilities. New planning systems are steadily moving from the era of DVH constraints applied as suggested by international recommendations and available on the commercial treatments planning systems (TPS) towards the definition of three-dimensional patient-specific constraints as part of the TPS optimization. Some of the commercial TPS are also able to take into account predictive model constraints. Accurate models for the prediction of toxicity allow for understanding of the complexity of dose-effect relationship which, within the context of personalized medicine, will help to devise tailored radiotherapy treatments. The nature of deformable organs whose planning dose is not representative of the real delivered dose may appear as a further issue.
