ESTRO 2021 Abstract Book

S71

ESTRO 2021

Conclusion Moderate DSC was demonstrated between observers with higher agreement for nodal outlines than primary tumour ones. Despite this IOV, the measured mean and % change in ADC between observers was not significantly different. RTT led evaluation of ADC using DW-MRI during RT for HNSCC is feasible. Refs: 1. Paterson C, et al (2017). Study of diffusion weighted MRI as a predictive biomarker...: The MeRInO study. ctRO: https://doi.org/10.1016/j.ctro.2016.12.003 . 2. Duffton A, et al (2019). OC-0415 MERINO study: Defining a standardised delineation method for repeated GTV assessment using DW MRI. Radiotherapy and Oncology.133:S215- S216.

Poster highlights: Poster Highlights 3: Outcome modelling

PH-0103 Outcome prediction for the prognosis of head and neck cancer patients based on deep learning J. Guo 1 , T. Zhai 2 , A. van der Schaaf 1 , R. J.H.M. Steenbakkers 1 , S. Both 1 , J. A. Langendijk 1 , P. M.A. van Ooijen 1 , N. M. Sijtsema 1 1 University Medical Center Groningen, Department of Radiation Oncology, Groningen, The Netherlands; 2 Cancer Hospital of Shantou University Medical College, Department of Radiation Oncology, Shantou, China Purpose or Objective Currently, prediction models employed in patient care are mostly based on clinical parameters. Imaging data, such as the pretreatment CT-scans, provide abundant information about on tumor characteristics and can be used to improve the outcome prediction. The objective of this work was to develop and test CT image based deep learning models for multi-outcome prediction of head and neck cancer (HNC) patients after definitive (chemo)radiation in combination with clinical parameters to improve the currently employed prediction models. Materials and Methods We experimented on a dataset of 444 HNC patients with contrast-enhanced CT-scans used for radiotherapy treatment planning. All patients underwent curative (chemo)radiation between 2007 and 2015. An experience radiation oncologist contoured the Gross Tumor Volume of all primary tumors. The dataset includes multiple clinical parameters, such as gender, age, TNM stage, clinical stage, treatment modality, WHO performance status and tumor site combined with HPV status. Multiple prognostic-related endpoints (local recurrence, regional recurrence, distant metastasis) were collected. We used the treatment time point 2017 to split the dataset into training (before 2017) with 234 patients and validation (after 2017) with 204 patients. We developed a multi-endpoint prediction model for the outcome of HNC patients. As illustrated in Figure 1, the model first takes as input 3D CT volumetric data that contain the delineated primary tumors to extract image features. The 3D image feature extraction was done through a 3D convolutional neural network. Then, the extracted image features were combined with the clinical parameters. Finally, a multi-layer perceptron was used to predict all the outcome endpoints at two years after treatment. All model parameters were trained end-to-end using the Adam optimizer with a learning rate of 0.001. The binary cross entropy loss for multiple outputs was used as the loss function.

Results