ESTRO 2024 - Abstract Book

S4426

Physics - Machine learning models and clinical applications

ESTRO 2024

674

Digital Poster

Automated right-sided breast planning via deep-learning left-sided model adaptation.

Michele Zeverino 1 , Wendy Jeanneret-Sozzi 2 , Francois Bochud 1 , Raphael Moeckli 1

1 Lausanne University Hospital, Institute of Radiation Physics, Lausanne, Switzerland. 2 Lausanne University Hospital, Radiation Oncology, Lausanne, Switzerland

Purpose/Objective:

To report the validation of an automated planning process for treating right-sided early-stage breast cancer under free-breathing conditions. This process was based on adapting the previously developed deep-learning (DL) model for left-sided deep inspiration breath hold treatments, which is currently in use at our clinic.

Material/Methods:

The automated process was developed in RayStation (RS) v12A TPS and applied to SIB early-breast cancer VMAT treatments delivering 48Gy/42.4Gy in 16 fractions for PTV_Boost and PTV_Breast, respectively.

In RS, the auto-planning optimization pipeline involves three fully automated sequential steps to generate the clinical dose, all based on the Model ROI association. Firstly, the U-NET model predicts the three-dimensional DL dose distribution based on the Model ROI association. Secondly, the predicted dose undergoes the so-defined “post processing”, where a voxel-based dose correction filter is applied to the predicted dose mostly regardless of the existing dose trade-offs between competing structures. Thirdly, the “post-processed” predicted dose is used as input for the “mimicking” process that involves the use of specific functions as well as classical dose objectives to produce the clinical dose during sequential cycles of iterations. Both the post-processing and the mimicking of the predicted dose are user-definable in the instruction file that comes along with the auto-planning model. This instruction file contains the sequence of the parameters assigned to each Model ROI used during the post-processing and the mimicking. Altering/removing the existing parameters or adding new ones has the effect of producing a different final solution starting from the same initial predicted dose.

The adapting strategy was initially tuned on an initial set of 8 patients with the goal of achieving at least the same quality of clinical manually designed plans that were taken as reference. It consisted of the following steps:

1. Run the existing left-sided DL model with the left-to-right swapping of the concerned structures in the Model ROI association as presented in Figure 1 to produce the DL-based predicted dose. 2. Evaluation of the predicted dose against the manual dose. Definition of a new post-processing parametrization for the predicted dose to approximate the post-processed predicted dose to the manual dose. 3. Evaluate the mimicked outcome of the adapted post-processed predicted dose and correct it when deviations from the manual plan occur, with a new parametrization of the mimicking process.