ESTRO 2023 - Abstract Book

S725

Monday 15 May 2023

ESTRO 2023

Figure 2 depicts DVHs of the accumulated dose distributions for two example patients, comparing daily adaptation and no adaptation for both dose weighting methods. For patient A, the dosimetric benefits of daily adaptation in the high-risk CTV diminish when moving from RBE-weighted to LET-weighted dose. For patient B, benefits in target coverage are still evident, but the prescription of 70 Gy in the high-risk CTV is missed.

Conclusion LET is critical for accurately calculating the biologically effective dose delivered by head-and-neck IMPT. The inclusion of LET in plan optimization should be considered to maximize the dosimetric benefits of daily adaptive proton therapy.

MO-0878 Dosimetric comparison of contouring techniques for adaptive radiotherapy A. Smolders 1,2 , T. Lomax 1,2 , D.C. Weber 1,3,4 , F. Albertini 1

1 Paul Scherrer Institute, Centre for Proton Therapy, Villigen, Switzerland; 2 ETH Zurich, Department of Physics, Zurich, Switzerland; 3 University Hospital Zurich, Department of Radiation Oncology, Zurich, Switzerland; 4 Inselspital, Bern University Hospital, University of Bern, Department of Radiation Oncology, Bern, Switzerland Purpose or Objective In online adaptive radiotherapy (ART), the treatment is reoptimized each fraction based on a daily image so that radiation can be delivered more precisely. Therefore, daily contours are needed, which can be obtained by automatic segmentation (AS) or image registration (IR). These un-corrected contours typically require manual adjustments, considered as a bottle neck for clinical introduction of ART. State of the art AS uses neural networks (NN) trained on a large dataset. Contrarily, IR relies on a single, but highly relevant example. In this work, we compare different methods for automatic contouring in ART by evaluating the dosimetric impact of using them for reoptimization in proton therapy. Materials and Methods The methods were evaluated on 5 non-small cell lung cancer patients (NSCLC) with a reference and 9 repeated CTs and 5 head and neck cancer (HNC) patients with 4-7 repeated CTs, all with manual contours of the relevant OARs and CTV defined by a radio-oncologist. We compared 5 techniques for automatic contouring in ART: rigid (RIR) and b-spline deformable IR (DIR), a commercial AS by LimbusAI (LC), pretrained NNs (PNN) and newly developed patient specific NNs (PSNN). The PNN were trained on a dataset with +- 110 images for the NSCLC patients and +- 350 images for HNC patients. The PSNN is similar to the PNN but was retrained in a second stage solely on the reference CT, so that the network overfitted on the patient under study. To assess the dosimetric impact of using automatic OAR contours, we reoptimized the plan on the daily CTs with automatic OAR and manual CTV contours. These plans were compared to the ones reoptimized on all manual contours using the DD2, the 2nd percentile of the voxel-wise dose differences, i.e. 98% of the volume receives a dose difference smaller than this value. We further evaluate the surface dice and the time to run the algorithm. To assess the impact of using automatic target contours, we reoptimized the plan on the daily CTs with automatic OARs and CTV. Only RIR, PSNN and DIR can contour the CTV. We compare the target coverage of the manual CTV for these plans, and benchmark them to the recalculated original treatment plan (as in non-adaptive therapy) and the plan reoptimized on all manual contours. Results The dose difference between the reoptimized plans is already small using non-perfect contours (RIR and PNN) and is even lower for PSNN and DIR (Tab. 1). PSNN is further much faster than DIR.