ESTRO 2024 - Abstract Book

S4520

Physics - Machine learning models and clinical applications

ESTRO 2024

1 Radiation physics, Department of Hematology, Oncology and Radiation Physics, Skåne University Hospital, Lund, Sweden. 2 Medical Radiation Physics, Department of Translational Medicine, Lund University, Malmö, Sweden. 3 Oncology, Department of Hematology, Skåne University Hospital, Lund, Sweden. 4 Oncology, Department of Clinical Sciences, Lund University, Lund, Sweden. 5 Department of Medical Physics, Memorial Sloan-Kettering Cancer Center, New York, USA. 6 Medical Radiation Physics, Department of Clinical Sciences Lund, Lund University, Lund, Sweden

Purpose/Objective:

In this study, we conducted a secondary analysis utilizing data derived from the randomized controlled phase III clinical trial ARTSCAN III [1] . The trial included patients with locoregionally advanced head and neck squamous cell carcinoma (HNSCC) and encompassed a cohort of 298 individuals recruited from 11 radiotherapy departments across Sweden. The focus of our secondary analysis centered on the late side-effects of xerostomia of grade II-III severity. Our analysis utilized a wide range of statistical models to comprehensively assess the risk of the late side-effect. Key factors included dosimetric variables, such as radiation dosage delivered to critical organs at risk (OAR) in conjunction with various clinical variables, all of which were essential to the modelling process. The aim of our study was to establish correlations between occurrence of late side-effects and parameters such as dose and clinical variables. In this study, a total of 298 patients with locoregionally advanced head and neck squamous cell cancer were selected for additional analysis. The patients were treated within a prospective randomized trial (ARTSCAN III). The patients were randomly assigned in a 1:1 ratio to receive cetuximab prior to commencing radiotherapy, followed by chemo radiation treatment. A secondary randomization was performed for patients with T3-T4 tumors, involving two radiation dosage regimens of either 68 Gy or 73.1 Gy, administered in 34 fractions. Following completion of radiotherapy, patients underwent regular follow-up assessments at three-month intervals during the initial two years, which were then extended to six-month intervals up to the fifth year. Of the initial patient cohort, 60 patients were excluded due to incomplete radiotherapy or missing data. The analysis incorporated diverse dosimetric parameters, including both dose (D) and volume (V) parameters, as well as patient demographic factors such as age, gender, N and T staging, HPV status, cetuximab (yes/no) prescribed dose and late side-effect outcomes. Feature selection from these parameters was carried out using both Univariate Logistic Regression and the ElasticNet machine learning method. The ElasticNet technique, which blends the L1 regularization (Lasso) and the L2 regularization (Ridge) methods, was employed to strike a balance between feature selection and regularization. This approach proves particularly valuable when working with datasets containing a large number of features. The chosen features were then integrated into both multivariate Logistic Regression and ElasticNet models. In addition to the statistical models a deep learning (DL) model was developed to incorporate three-dimensional (3D) dose information for the purpose of correlating spatial dose distributions with imaging data. A custom 3D ResNet architecture was developed. The model's inputs consisted of 3D dose distributions, CT scans, and contours of OARs. Material/Methods:

Results: