[Poster] EPID-based in vivo transit dosimetry in external beam radiotherapy: prediction of portal dose images using artificial neural networks

Côme Mével Dutertrea^b, Alexandre Hakimi^a, Eric Fadel^a, François Smekens^a, Xavier Franceries^b, François Husson^a

Introduction: EPID-based in vivo transit dosimetry using the forward technique requires the accurate calculation of the predicted (reference) image to be compared in terms of absolute dose with the acquired image. The Collapsed Cone Convolution (CCC) algorithm is a fast method for accurate dose predictions suitable to transit dosimetry, and allows a distinction between the dose generating phenomena. The Detector Total Dose (DTD) at the computation plane is the sum of the Attenuator Transmitted Dose (ATD) from direct photons beam transmission and the Attenuator Scattered Dose (ASD) from beam interaction in the attenuator. This work introduces a Machine Learning workflow based on CCC ground-truths, using a Convolutional Neural Network (CNN) for ATD calculation and a conditional Generative Adversarial Network (cGAN) for ASD computation.

Conclusion: The full ML model allows: ● CCC equivalent predictions in terms of total dose at the EPID level ● Lower computation time for identical CPU/GPU (~100x less) ● Generalization across various tumoral sites ● Restitution of physics phenomena (transmission, scattering)