Explainable symbolic regression for tumor growth prediction of paragangliomas

Abstract

Head and neck paragangliomas are generally benign tumors that often grow slowly or not at all. MRI surveillance is common, but growth may cause irreversible symptoms, while treatment also carries risks. Predicting growth could inform treatment decisions.

This thesis investigates tumor growth measurement and prediction. Automated MRI segmentation performed at least as well as human observers, enabling longitudinal data collection from 311 tumors, with a median of five measurements each—substantially more than in previous studies.

Symbolic regression was extended to discover growth function structures beyond predefined equations. This approach identifies general function classes whose coefficients can be fitted to each tumor, allowing shared mathematical structures to describe individual growth trajectories. Another method was developed to identify multiple high-quality, meaningfully different models, enabling the discovery of alternative growth function classes rather than a single solution. Three relevant growth patterns emerged, with Gompertz-like, S-shaped functions performing best. Finally, methods for estimating prediction uncertainty were compared, considering statistical accuracy and practical usefulness in distinguishing stable from growing tumors. Since uncertainty depends on the number of measurements and the prediction horizon, quantifying it is essential for assessing the reliability of individual predictions. Together, these methods provide a foundation for patient-specific forecasts that could support decisions between surveillance and treatment.