ESTRO 2025 - Abstract Book

S1026

Clinical – Head & neck

ESTRO 2025

2447

Poster Discussion MRI far exceeded CT/PET and clinical benchmark model for local recurrence prediction in oropharyngeal cancer: a deep learning-based investigation Baoqiang Ma 1 , Jiapan Guo 1,2,3 , Lisanne V. van Dijk 1,4 , Johannes A. Langendijk 1 , Peter M.A. van Ooijen 1,2 , Stefan Both 1 , Nanna M. Sijtsema 1 1 Radiation Oncology, University Medical Center Groningen, Groningen, Netherlands. 2 Machine Learning Lab, Data Science Center in Health, University Medical Center Groningen, Groningen, Netherlands. 3 Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence, University of Groningen, Groningen, Netherlands. 4 Radiation Oncology, University of Texas MD Anderson Cancer Center, Houston, USA Purpose/Objective: Due to the superior soft tissue contrast to CT, MRI may contain more information for outcome prediction of oropharyngeal cancer (OPC). However, MRI's signal variability complicates quantitative analysis. Convolutional Neural network (CNNs) show promise in improving outcome prediction by extracting more robust features. This study aims to develop CNNs for the prediction of local control (LC), regional control (RC) and overall survival (OS) using T1 and T2 MR-images and compare their predictive performance with a clinical benchmark model and CT and PET-based models in OPC patients treated with (chemo)radiotherapy.

Material/Methods:

A dataset comprising 266 OPC patients who received (chemo)radiotherapy was assembled. Each patient’s data included a pretreatment axial T1-weighted scan (T1) and a coronal T2-weighted scan (T2) (resolution around 0.5x0.5x3 mm3), registered CT and PET scan (resolution: 1x1x2 mm3), contoured Gross Tumor Volume (GTV) of primary tumor, clinical parameters and follow-up information regarding LC, RC and OS. Patients were divided into a training set (n = 186) and a test set (n = 80). Various 2D ResNet18 [1] models were trained using either T1 or T2 inputs of the contoured GTV volume with and without a margin for outcome prediction (illustrated in Figure 1). Additionally, several 3D ResNet18 models were trained, each using only one imaging modality (either T1, T2, CT, or PET) within the same GTV