Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

A Machine Learning Classifier for Predicting Recurrence in Oropharyngeal Cancer

T Chinnery*, P Lang, A Nichols, S Mattonen, Western University, London, ON


PO-GePV-M-42 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: To develop a machine learning classifier integrating radiomic image features and clinical data to predict recurrence in oropharyngeal cancer.

Methods: A dataset of patients (n=295) with oropharyngeal cancer (OPC) treated with chemoradiation at the London Regional Cancer Program was used for this study. Primary tumor and nodal volumes were contoured on pre-treatment planning CT images. The Quantitative Image Feature Engine was used to compute size, shape, first-order intensity, and second-order textural radiomic features from these volumes of interest. The dataset was split into independent training (n=206) and testing (n=89) datasets. Feature selection was applied to select the optimal features to predict recurrence and a support vector machine (SVM) classifier was built using the selected features on the training dataset. The SVM’s performance was assessed in the testing dataset based on the metric of the AUC. A baseline model comprised of only clinical features was compared to radiomics.

Results: A total of 63 patients (21%) developed a recurrence. Through feature selection, 20 predictive features were selected. This included one first-order intensity feature, seven textural, nine shape, and three clinical features (HPV status, smoking status, nodal stage), among both the primary tumor and nodal volumes. The SVM model achieved an AUC on the testing dataset of 0.66 [95% CI: 0.53-0.77], in comparison to an AUC of 0.49 [95% CI: 0.36-0.62] for the model comprised of clinical features alone (p=0.06).

Conclusion: Radiomic features augmented traditional clinical features for predicting recurrence in our relatively small testing dataset. To further improve predictive performance, we plan to implement additional clinical data, as well as investigate other machine learning models. Once refined and validated, this model could assist in identifying patients at a higher risk of recurrence, who may benefit from more personalized treatment options.



    CAD, Quantitative Imaging, CT


    TH- Response Assessment: Radiomics/texture/feature-based response assessment

    Contact Email