ESTRO 2023 - Abstract Book

S1353

Digital Posters

ESTRO 2023

Results Figure 1 shows the mean DVH of the ipsilateral-lung and heart (over all 16 patients) within the SD variability. Inter-institute SD of the DVH prediction in the dose range from 20% to 80% was 1.7% and 1.2% for the ipsilateral lung and the heart, respectively. The inter-institute variability in terms of Dmean was 0.5 Gy and 0.3 Gy for the ipsilateral lung and the heart, respectively; while for contralateral OARs the SD was <0.2 Gy. Figure 2 shows the distribution of the mean predicted heart and ipsilateral lung dose for every institute through all patients of the test set. Average institutional values were between 1.1 and 2.3 Gy for the heart and between 4.3 and 5.9 Gy for the ipsilateral lung. Models showed high transferability in terms of PC1 for both in-field OARs: the value of PC1 for each model in the test set was within 90th percentile of the corresponding training set in more than 80% of cases (80-100%), suggesting no relevant differences among models in terms of transferability.

Conclusion Results show limited inter-institute variability of plan prediction KB models, and high model’s transferability. These findings encourage the building of a robust benchmark model, with large potentials for plan QA, audit, education/tutoring and large-scale KB plan automation.

This study is supported by an AIRC grant (IG23150).

PO-1653 Auto-contouring of cardiac avoidance region for cardiac sparing lung radiotherapy

T. Marchant 1,2 , G. Price 2 , A. McWilliam 2 , E. Henderson 2 , D. McSweeney 2 , K. Banfill 2,3 , J. King 3 , C. Barker 3 , C. Faivre-Finn 2,3

1 The Christie NHS Foundation Trust, Christie Medical Physics & Engineering, Manchester, United Kingdom; 2 The University of Manchester, Division of Cancer Sciences, Manchester, United Kingdom; 3 The Christie NHS Foundation Trust, ., Manchester, United Kingdom Purpose or Objective Heart dose during lung radiotherapy is increasingly recognised to have a significant effect on patient outcome. Several studies have identified an area at the base of the heart where dose is strongly correlated with overall survival. To introduce a dose limit for this cardiac avoidance area (CAA) it first needs to be contoured for treatment optimisation. An auto contouring solution for this region will reduce the time spent contouring by clinicians, improve contouring consistency, and make implementation of heart sparing radiotherapy feasible. In this abstract we report the development and validation of a deep-learning based auto-contouring method for the proposed CAA. Materials and Methods The CAA chosen at our institution includes the right atrium, aortic valve root, and proximal portions of the left and right coronary arteries. An existing fast, efficient 3D Convolutional Neural Network [DOI: 10.1016/j.media.2022.102616] was adapted for cardiac substructures. The model is designed for training with a small dataset, here 42 4DCT-AVG scans of patients with lung cancer with manual contours for CAA and associated subregions. When trained, contours are produced using CPU only, facilitating easy deployment of the model without specialist hardware. An initial step identifies the heart location in the CT and crops the image around it [DOI: 10.48550/arXiv.2209.06042]. Contouring runs on the cropped image to infer masks representing each cardiac substructure, before these are combined into the final CAA contour. Auto-contour quality was assessed by geometric comparison to manual contours for 10 patients not used for training. Further, auto contouring robustness was evaluated using 198 CT scans. Forty-five randomly selected cases were rated by an expert