CHARACTERIZATION OF BREAST RT TREATMENT BY MULTI-INSTITUTIONAL AI-KNOWLEDGE-BASED PLANNING OPTIMIZATION MODELS

This thesis develops a decision-making tool based on artificial intelligence to predict dose distribution to Organs at Risk, particularly the heart, in breast cancer radiotherapy. The goal is to enable physicists to quickly assess, via computed tomography (CT), whether a tangential field plan is adequate or if alternative techniques, such as breath-hold CT or Volumetric Modulated Arc Therapy (VMAT), are necessary to optimize therapy and avoid unacceptable doses to the heart. The work is part of the AIRC MIKAPOCO, which has developed multi-institution benchmark models for automatic planning of tangential fields. Using RapidPlan® software, the project aimed to create a unified multicenter model, reducing planning variability. However, no tool exists to quickly determine, based solely on CT and without segmentations, whether other techniques may be more appropriate. This tool is crucial for optimizing resources and customizing treatments, especially in the case of breath-hold, which requires patient selection, as it is complex to implement and not justified for all patients. First, an automatic segmentation model of the CTV was developed using a U-Net neural network on the MIKAPOCo dataset, achieving good agreement among the centers, comparable to manual inter-variability. Then dose prediction was conducted elegantly, treating isodoses as structures and using the same network designed for CTV segmentation, deriving the cardiac volume receiving 20 Gy and 36 Gy. Furthermore, in line with the aim of providing a tool for precise characterization of the treatment process, predictive models for skin toxicity and clinical outcomes were developed. For late skin toxicities, an association emerged between irradiated skin volumes and the risk of side effects, with specific constraints helping to control the risk. Regarding clinical outcomes, a densitometric analysis of the CTV revealed the protective role of adipose tissue on endpoints such as Local Relapse Free Survival, Distant Relapse Free Survival, and Overall Survival.