• AI
• DL
• EPID
• IVD
• Phantoms
• Radotherapy

Modern radiotherapy requires high precision in dose delivery to ensure tumor control and minimize damage to healthy tissue. In-vivo dose monitoring is crucial due to treatment complexity and potential errors. EPIDs (Electronic Portal Imaging Devices) offer real-time data and, recently, a Deep Learning model has been developed to convert EPID images into clinically relevant Portal Dose (PD) predictions. Trained on a large dataset of EPID images and dose distributions from the Monaco TPS, the model is evaluated in this thesis for two main goals: assessing its applicability in an in-vivo alert system and quantifying its sensitivity to treatment errors. The alert system aims to detect significant discrepancies between predicted and delivered doses in real-time. Validation involved EPID images of various phantoms (anthropomorphic, modular, solid water) with controlled variations in Monitor Units (MU), positioning, gantry angle, and added thickness. The model showed good sensitivity: GPR (3%, 3 mm) dropped below 95% for MU >102.4–109.3, positioning errors >3.9–4.6 mm, gantry deviations >8.9°, and added thicknesses of 1–3 mm. These results highlight the importance of evaluation metrics and guide optimization for clinical implementation. Further improvements to the model and phantom design were also identified.