ESTRO 2024 - Abstract Book

S5112

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

2251

Digital Poster

NTCP modeling for severe oral mucositis from HNC radiotherapy using radiomics and machine learning

Kanyapat Buasawat 1 , Sasikarn Chamchod 1,2 , Todsaporn Fuangrod 1 , Thiansin Liamsuwan 1

1 Chulabhorn Royal Academy, Princess Srisavangavadhana College of Medicine, Bangkok, Thailand. 2 Chulabhorn Royal Academy, Radiation Oncology Department, Chulabhorn Hospital, Bangkok, Thailand

Purpose/Objective:

To develop machine learning-based normal tissue complication probability (NTCP) models for severe oral mucositis (grade 3 or worse) for head and neck cancer (HNC) patients treated with volumetric modulated arc radiation therapy (VMAT) and intensity modulated radiation therapy (IMRT) using clinical, dosimetric and radiomic features.

Material/Methods:

A cohort of 200 HNC patients were included in this study. A total of 14 clinical features, 38 dosimetric features and 214 radiomic features of the oral cavity and the parotid gland were initially acquired. The radiomic features have been tested for robustness in term of interobserver delineation variability. The Pearson correlation test was employed to reduce the feature redundancy. Four feature selection algorithms, including minimum redundancy maximum relevance (MRMR), model-based random forest (MBRF), Relief and sequential, as well as 4 machine-learning based classification algorithms, including logistic regression (LR), support vector machine (SVM), random forest (RF) and k nearest neighbor (k-NN) were used for the modeling. The patient dataset was split into 80% training and 20% testing. The five-fold cross validation was applied on the training dataset for hyperparameter tuning and to find the best set of features for each combination of feature selection and classification algorithms. The model performance was evaluated in term of the area under the receiver operating characteristic curve (AUC).

Results:

The combination of SVM and MBRF with 24 selected features was found to give the best prediction with the AUC of 0.82, followed by the combinations of LR and MBRF (25 features, AUC = 0.80), k-NN and MRMR (25 features, AUC = 0.76), and RF and MBRF (7 features, AUC = 0.75). Radiomic and dosimetric features were among the majority of the selected features for the best model (SVM with MBRF).

Conclusion:

The machine-learning based models investigated in this work showed good performance in predicting severe oral mucositis, with the MBRF feature selection algorithm together with the SVM classification algorithm outperformed the other investigated models.