ESTRO 2024 - Abstract Book

S4517

Physics - Machine learning models and clinical applications

ESTRO 2024

2013

Digital Poster

External validation of a deep-learning prediction model for mandibular osteoradionecrosis

Laia Humbert-Vidan 1,2 , Andrew P King 3 , Vinod Patel 4 , Christian R Hansen 5 , Jørgen Johansen 6 , Teresa Guerrero Urbano 7

1 Guy's and St Thomas' NHS Foundation Trust, Department of Medical Physics, London, United Kingdom. 2 King's College London, School of Cancer and Pharmaceutical Sciences, London, United Kingdom. 3 King's College London, School of Biomedical Engineering and Imaging Sciences, London, United Kingdom. 4 Guy's and St Thomas' NHS Foundation Trust, Department of Oral Surgery, London, United Kingdom. 5 University of Southern Denmark, Department of Clinical Research, Odense, Denmark. 6 Odense University Hospital, Department of Oncology, Odense, Denmark. 7 Guy's and St Thomas' NHS Foundation Trust, Department of Clinical Oncology, London, United Kingdom

Purpose/Objective:

Mandibular osteoradionecrosis (ORN) is a severe late side effect affecting patients undergoing radiation therapy (RT) for head and neck cancer. The vascularisation, density and composition of the bone varies across the mandible. Thus, some areas are more prone to ORN than others. Additionally, the radiation dose distribution in the mandible is not homogeneous. Existing ORN studies are based on dose-volume (DVH) parameters that do not capture this clinically relevant spatial information. The use of radiation dose distribution maps as the dose information for predicting mandibular ORN has recently been introduced [1, 2]. However, these works featured only internal validation on a holdout subset of the data used for training. Model validation studies in the context of spatial dose NTCP modelling are limited, probably due to the technical complexities involved in the data preparation process. However, external validation of such models is necessary for their acceptance in a clinical context largely dominated by DVH-based models. The current study aimed to externally validate an existing DL-based ORN prediction model [1].

Material/Methods:

A 3D DenseNet-40 (DN40) convolutional neural network (CNN) was trained on 3D radiation dose distribution maps of the mandible for the binary classification of ORN vs. no ORN subjects. The model was developed and internally validated in a UK population (92 ORN cases and 92 controls) and was externally validated in an independent dataset from a Danish population (41 ORN cases and 41 controls). In both validation processes, the model’s performance was assessed in terms of its discriminative ability and calibration. Platt scaling was applied to improve the calibration of the predicted probabilities at external validation. Finally, the DN40 model discriminative performance on the external dataset was compared to that of a Random Forest model on the corresponding DVH data.

Results: