ESTRO 2024 - Abstract Book

S3543

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

ESTRO 2024

1274

Poster Discussion

Inter-institute transferability of KB plan prediction models for L-WBI on DIBH candidates.

Roberta Castriconi 1 , Alessia Tudda 1 , Giovanna Benecchi 2 , Francesca Dusi 3 , Valeria Landoni 4 , Aldo Mazzilli 2 , Caterina Oliviero 5 , Alessandro Scaggion 3 , Antonella del Vecchio 1 , Lorenzo Placidi 6 , Claudio Fiorino 1 1 IRCCS San Raffaele Scientific Institute, Medical Physics, Milan, Italy. 2 University Hospital of Parma AOUP, Medical Physics, Parma, Italy. 3 Veneto Institute of Oncology IOV–IRCCS, Medical Physics, Padua, Italy. 4 IRCCS Istituto Nazionale dei Tumori Regina Elena, Medical Physics, Rome, Italy. 5 University Hospital, “Federico II”, Radiotherapy, Naples, Italy. 6 Fondazione Policlinico Universitario A. Gemelli IRCCS, uosd medical physics and radioprotection, Rome, Italy

Purpose/Objective:

To test if Knowledge-based (KB) plan prediction models trained in eight institutes based on free-breathing (FB) left whole-breast irradiation (L-WBI) were transferable to patients whose planning CT was obtained in deep inspiration breath-hold (DIBH) modality. Second aim was to quantify inter-institute variability of additional 4 KB institutional models trained on DIBH patient’s cohorts.

Material/Methods:

KB plan prediction models for L-WB using tangential-fields (TF) were generated within a multi-centric study: 8 institutions developed their FB-KB models (INST1-INST6; INST9-INST10) by using RapidPlan (Varian Inc.) following previously agreed guidelines for modeling and validation. Four out of eight institutes delivering DIBH developed additional KB models (INST1, INST2, INST6, INST10) using data referred to their DIBH patients, following the same common methodology used to train KB models. In the first part of the work, we performed a cross validation test with 20 new DIBH randomly chosen patients from five out of the eight institutes (INST1, INST2, INST4, INST6, INST10): each model was run on this test cohort. Then, cross validations were also performed using the same 20 DIBH test set patients running the four KB-DIBH models. The transferability of FB-KB models and DIBH-KB models on DIBH patients was assessed by considering the overlap of the geometric Principal Component score 1 (PC1) when applied to the test patients. For the transferability analysis of KB-DIBH models, test patients from the same DIBH-KB model institute were not included. We considered both FB and DIBH KB models from an institute “A” (INSTA-FB/INSTA-DIBH model) transferable on a DIBH-CT of an institute “B” patient (INSTB-DIBH) when the ipsilateral lung and heart PC1 was within the 10th- 90th percentile of the training set of each KB-FB/KB-DIBH model. Dosimetric parameters were analyzed. In addition to heart and lung mean predicted dose, we also included lung V20Gy, heart V5Gy and V25Gy with the aim of quantifying inter-institute variability in DVH prediction.

Results:

Figure 1 A) and B) show the heart and ipsilateral lung mean predicted dose of FB-KB models on DIBH candidates, with SDmean=0.4Gy for heart and SDmean=0.8 Gy for ipsilateral lung. Inter-institute SD of DVH prediction in the dose range from 20% to 80% was 0.4% and 2% for heart and ipsilateral lung respectively. Regarding heart SDV5Gy=1.4% and SDV25Gy=0.8% were found, while SDV20Gy=2.8% resulted for ipsilateral lung. Figure 1C and Figure 1D shows the fraction of DIBH patients with PC1 within the 90th percentile of the training set for all FB-KB models. Concerning heart, INST1DIBH’s model heart PC1 was within 90th for all patients, while only 3 out of 8 FB-