ESTRO 2024 - Abstract Book

S3579

Physics - Dose prediction, optimisation and applications of photon and electron planning

ESTRO 2024

1 IRCCS San Raffaele Scientific Institute, Medical Physics, Milan, Italy. 2 Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Medical Physics, Rome, Italy. 3 Azienda USL-IRCCS, Medical Physics Unit, Department of Advanced Technology, Reggio Emilia, Italy. 4 Fondazione IRCCS Istituto Nazionale dei Tumori (INT), Progetto Prostata, Milan, Italy. 5 Veneto Institute of Oncology IOV–IRCCS, Medical Physics Department, Padova, Italy. 6 SOC Fisica Sanitaria, ASU FC, Udine, Italy. 7 Istituto Nazionale dei Tumori Regina Elena, Medical Physics Department, Roma, Italy. 8 Fatebenefratelli Isola Tiberina – Gemelli Isola, UOC di Radioterapia Oncologica, Roma, Italy. 9 University Hospital of Parma AOUP, Medical Physics Dept, Parma, Italy. 10 Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Medical Physics Dept, Rome, Italy. 11 Federico II University Hospital, Medical Physics, Napoli, Italy

Purpose/Objective:

To generate and validate two Knowledge Based (KB) models obtained merging clinical plans from several Italian centers for right and left whole breast irradiation (WBI) using tangential fields (TF, wedged/field-in-field). Previous evaluations of inter-institute variability and assessment of model’s transferability were accomplished. The resulting KB models were intended as benchmarks for potentially large-scale implementation.

Material/Methods:

Ten Institutes (INST1-INST10) developed their KB models for WBI. Methods to generate the models, including minimum number of patients (>70), geometry definition, outlier identification/elimination criteria and internal validation were previously agreed and strictly followed by each participating institute in set-up their models. Ten/eight KB-institute models for right/left breast respectively, were generated (1481 total patient) and imported into a dedicated research station (Eclipse v16.1). The distribution of anatomical (PTV/OARs volumes) and dosimetric parameters (PTV V95%, D99%, V105%,D1%) related to the original test sets were first quantified and prediction bands of OARs (heart, contralateral and ipsilateral lung, contralateral breast) were extracted on 36 test patients of the same institute, not included in KB models, through a dedicated script. Inter-institute variability was quantified comparing SD of predicted DVHs and Dmean of OARs across different models. We assessed the transferability of KB models among centers by reporting the overlap of the geometric Principal Component (PC1) of each model when applied to the test patients of the other institutes: we considered model INSTA as transferable to INSTB if the OARs PC1 values of the INSTA model within the 10 th -90 th percentile of the training set of INSTB in more than 80% of the cases. Two KB-Benchmark models for right and left breast respectively were obtained using data (849 patients) from available institutes generating transferable KB models . After tuning of the models, we performed a cross validation test running KB-Benchmark models on an external patient set (5 from each center): bands of predictions for the OARs were extracted and compared to the ones of the institutional KB-institute.

Results:

Concerning KB-institute models analysis, inter-institute variability was 1.8% for ipsilateral lung for both right and left breast models and 1.6% for heart for the left breast model (in the 20-80% dose range). Concerning right breast, the predicted ipsilateral lung Dmean test set patients were correlated with median PTV D99% data set (R2= 0.78), once excluded INST6 (due to a different contouring approach); PC1 values were always within the range of applicability for each model (> 94% of patients) apart INST6, that was then excluded from the benchmark model training. Concerning left breast, we found high transferability for both ipsilateral lung and heart (> 80% of success) among institutes; Dmean heart prediction was consistent among institutes (1.8 Gy) excluding INST1, that was then excluded from the benchmark model training. Full data of six (456 pts) and five (393 pts) consistent institutional models were available for generating right and left breast Benchmark models respectively.