ESTRO 2024 - Abstract Book

S4470

Physics - Machine learning models and clinical applications

ESTRO 2024

[3] S. Graham, Q. D. Vu und S. E. A. Raza, HoVer-Net: Simultaneous Segmentation and Classification of Nuclei in Multi Tissue Histology Images, arXiv, 2019.

[4] J. N. Kather, N. Halama und A. Marx, 100,000 histological images of human colorectal cancer and healthy tissue, Zenodo, 2018.

[5] S. Graham, M. Jahanifar und A. Azam, Lizard: A Large-Scale Dataset for Colonic Nuclear Instance Segmentation and Classification, arXiv, 2021.

1333

Digital Poster

Predicting late taste loss with deep learning NTCP model using 3D dose, CT and OAR segmentations

Hendrike Neh, Suzanne P. M. de Vette, Hung Chu, Peter M. A. van Ooijen, Nanna M. Sijtsema, Johannes A. Langendijk, Lisanne V. van Dijk

UMCG, Radiotherapy, Groningen, Netherlands

Purpose/Objective:

Taste loss is a common yet complex toxicity of head and neck cancer (HNC) radiotherapy affecting more than 25% of patients at 6 months after treatment. Normal tissue complication probability (NTCP) models can predict radiation induced toxicities, such as taste loss, and typically utilize clinical variables and single dose-volume parameters from selected organs-at-risks (e.g., mean oral cavity dose). Adequate toxicity risk predictions can guide treatment planning to minimize taste loss, and subsequently improve quality of life for HNC survivors. Recent developments in artificial intelligence (AI) allow for the possibility to develop deep learning models that can predict radiation-induced toxicity based on the full 3D information without limiting this to discrete clinical and 1D dose-volume parameters. This study aims to improve NTCP model performance for taste loss at 6 months after treatment in a large HNC patient cohort by exploiting the entire 3D dose distribution, CT and organs-at-risk segmentations.

Material/Methods:

The patient cohort included 949 HNC patients treated with curative intent between 2007 and 2022. The endpoint is moderate-to-severe patient-rated taste loss at 6 months corresponding to the two highest scores on the EORTC QLQ H&N35 4-point Likert scale (none, mild, moderate and severe). For model development, the patient cohort was split into a training (70%), validation (15%) and unseen test set (15%). Deep learning model input consisted of the 3D dose distribution, planning CT and all common OAR segmentations of the head and neck region. The Residual deep convolutional network (Resnet-DCNN) with four convolutional blocks was chosen for this application (Figure 1). Internal validation was performed with a 5-fold cross-validation to ensure robustness. Hyperparameter tuning was aided by the Optuna framework. The previously published late taste loss NTCP model trained on 750 HNC patients (Van den Bosch et al., 2021) with the predictors oral cavity mean dose, parotid gland mean dose and age was used as