ESTRO 2024 - Abstract Book

S4542

Physics - Machine learning models and clinical applications

ESTRO 2024

Alessandro Cicchetti 1,2 , Monica Vincenzi 1 , Roberta Castriconi 1 , Paola Mangili 1 , Andrei Fodor 3 , Maria Giulia Ubeira Gabellini 1 , Anna Chiara 3 , Chiara Deantoni 3 , Martina Midulla 3 , Martina Mori 1 , Marcella Pasetti 3 , Gabriele Palazzo 1 , Roberta Tummineri 3 , Miriam Torrisi 3 , Tiziana Rancati 2 , Antonella del Vecchio 1 , Nadia G Di Muzio 3 , Claudio Fiorino 1 1 IRCCS San Raffaele Scientific Institute, Medical Physics, Milan, Italy. 2 Fondazione IRCCS Istituto Nazionale dei Tumori di Milano, Data Science Unit, Milan, Italy. 3 IRCCS San Raffaele Scientific Institute, Radiotherapy, Milan, Italy

Purpose/Objective:

Acute skin reactions are common side effects of breast Radiotherapy (RT). An NTCP model for predicting the onset of such toxicity by combining skin DVH and clinical factors was previously developed on a large mono-institutional group of patients treated with 3d-CRT using tangential wedged or field-in-field (FiF) techniques in 2009-2017.

Here, we present a temporal and technical validation analysing a new cohort of patients treated in 2017-2021 with the inclusion of advanced delivery techniques.

Material/Methods:

Clinical and dosimetry data of 1820 patients treated with 40 Gy in 15 fractions without additional boost were considered for the entire study. The model training was developed on 1128 patients (2009-2017) treated using manually optimised tangential techniques (3d-CRT, wedged/ FiF). We analysed 692 patients (2017-2021) for the validation study. The treatment was delivered using static and modern techniques, including volumetric and intensity-modulated arc approaches: VMAT, ViTAT(Virtual Tangential Arc Technique) and TomoDirectR (delivered with Tomotherapy machines). Particularly, 41.6% of patients were treated using arc techniques (46 ViTAT and 242 VMAT), and the remaining 58.4% with static techniques (299 3d-CRT and 105 TomoDirect, TD). Starting from the planning CT, a 3D “skin” structure was automatically defined as the 5 mm inner expansion from the outer contour of the Body representing the derma layers. We delineated and extracted absolute dose-volume histograms (DVH) using MIM software and MIM assistant. Skin toxicity was assessed during follow-up (FU) using the RTOG (Radiation Therapy Oncology Group) scale criteria. Patients who developed moderate-severe effects (ACUG2+) within 15 days from the end of RT were considered as the ones experiencing toxicity. The symptoms scoring in the two cohorts were assessed at different times, and the doctors accomplishing this task were the same in part. The association of 27 clinical features and 15 dosimetric variables was performed in the development study. After a preprocessing based on variable p-value, multivariate models were tested through backward variable selection. We used the Area under the ROC curve (AUC) and the Hosmer-Lemeshow test to assess the predictive performance. The resulting model was then tested in the validation group through calibration plots: the slope of the resulting linear fit between predicted and expected toxicity rates, R2, p-value and intercept were reported. The validation was performed (i) in the whole validation cohort and separately in the subgroups of patients treated with (ii) tangential like techniques (3d-CRT and TD, n=404) and (iii) arc techniques (VMAT/ViTAT, n=288)

Results: