Deep learning methods for 2D in-vivo dose reconstruction with EPID detector

Over the past two decades, radiotherapy has seen a steep increase in technological complexity of treatment preparation and treatment execution, calling for the development of more and better quality assurance tools and procedures. In-vivo dosimetry has emerged as a very powerful tool for treatment verification in conjunction with Electronic Portal Imaging Devices (EPIDs). However, EPIDs present several drawbacks like non-water equivalence and cumbersome calibration procedures. Artificial intelligence, specifically Deep Learning (DL), plays a crucial role in this context. A critical step is modeling the EPID response to estimate 2D dose distributions (Portal Dose, PD). This work introduces a DL-based methodology aimed at converting EPID responses into PD images. The proposed procedure is fully data driven, being based on advanced DL methods. This approach will allow bypassing the complex calibration steps needed to overcome the non-water equivalent panel.