ESTRO 2024 - Abstract Book

S4437

Physics - Machine learning models and clinical applications

ESTRO 2024

Automated treatment planning can be achieved by using ML methods alone but also by combining ML with optimization models and algorithms. In this work an ensemble approach joining RL, optimization and fuzzy inference systems is presented for fully automated treatment planning without resorting to large training datasets.

Keywords: Reinforcement Learning, optimization, IMRT

References:

[1] J. Dias, H. Rocha, T. Ventura, B. Ferreira, e M. C. Lopes, «Automated fluence map optimization based on fuzzy inference systems», Medical physics, vol. 43, n.o 3, pp. 1083–1095, 2016.

[2] R. S. Sutton e A. G. Barto, Reinforcement learning: An introduction. MIT press, 2018.

764

Proffered Paper

DIVE-ART: towards Dosimetrically Informed Volume Editions of automatically segmented volumes

Benjamin Roberfroid 1 , Edmond Sterpin 1,2,3 , John A Lee 1 , Xavier Geets 1,4 , Ana M. Barragán-Montero 1

1 UCLouvain, IREC-MIRO, Brussels, Belgium. 2 KU Leuven, Department of Oncology, Laboratory of Experimental Radiotherapy, Leuven, Belgium. 3 Particle Therapy Interuniversity Center Leuven, PARTICLE, Leuven, Belgium. 4 Cliniques universitaires Saint-Luc, Department of Radiation Oncology, Brussels, Belgium

Purpose/Objective:

Automatic segmentation in radiation therapy has raised considerable interest for many years, since it allows the clinicians to automate the tedious task of manually contouring anatomical volumes. The accuracy of these algorithms is crucial, particularly in the context of online adaptive radiotherapy, where the time for review and correction should be as short as possible. However, the lack of quality assurance tools requires visual inspection and manual correction of auto-segmented contours. In this work, we propose to reduce these manual operations thanks to a 3D heatmap indicating the regions where the corrections are estimated to have hardly any dosimetric impact: the DIVE (Dosimetrically Informed Volume Edition) map.

Material/Methods:

The DIVE-map for a given organ is generated as follows:

1. First, the auto-segmented organ is deformed following different amplitudes and directions, resulting in N deformation scenarios (S d,d=1,..N, N=162).