Survival Analysis Using Deep Learning Models for Head and Neck Cancer Patients
Abstract
This thesis made deep learning models for time-to-event prediction on patients with head and neck cancer. The follow-up period was split into ten time intervals, spanning a five year period. The models predicted the time until two different endpoints, overall survival and disease free survival. Models, based on the EfficientNet architecture, were given combinations of CT and PET images, and contours around the primary tumor and nodal areas, as input. The models were evaluated using Harrell's Concordance Index (C-index), the Area Under the Receiver Operating Characteristic Curve (AUC) and the Integrated Brier Score (IBS) as metrics, using one internal dataset to train and validate the models, and one external dataset for testing. The model that achieved the highest overall performance utilized CT, PET, and a primary tumor contour as inputs, and was able to achieve a C-index of 0.74, AUC of 0.69 and IBS of 0.16 on the internal dataset. The predictions on the overall survival endpoint were generally of a higher score than predictions on the disease free survival endpoint, across all metrics.
Additionally, the models were assessed on their explainability, detailing how the model predictions related to the observed real data, and what parts of the images the models used for predictions. This was done using Kaplan-Meier curves, and saliency maps from the Variance of the Model Gradients method and the Shapley Additive Explanations method. The Kaplan-Meier curves indicated that the models generally overestimated the survival probabilities of all patients. Saliency maps generated using the Variance of the Model Gradients method and the Shapley Additive Explanations method showed that the PET modality was the most influential in model predictions, while the CT modality had the least influence. The models mostly took information from the primary tumor area when predicting on the overall survival endpoint, and from both the primary tumor and nodal areas when predicting on the disease free survival endpoint. The Shapley Additive Explanations method for explaining the model predictions proved to show the same areas as the Variance of the Model Gradients method.