NIMG-02. DEVELOPING A RADIOMIC HIERARCHICAL GAUSSIAN PROCESS BOOSTING MODEL TO PREDICT PRIMARY TUMOR ORIGIN FROM MULTICENTRIC LONGITUDINAL MRI DATA OF BRAIN METASTASES

Abstract Focal neurologic deficits due to brain metastases might be the initial presentation in patients with cancer. Common approaches to investigate the occult primary malignancy include a whole-body PET-CT, whole-body MRI, studying tumor markers, or histopathological examination. Most of these modalities are expensive, time-consuming, invasive, and delay treatment. The field of radiomics leverages quantitative features extracted from images to improve the decision-making process in clinical practices. Although not easily understandable, these features are proxies for the tumor’s shape, contour, texture, and intensity patterns. They can help understand tumor heterogeneity, predict tumor behavior, and assess treatment response. We propose a robust radiomic workflow using longitudinal MRI data from 914 high-resolution imaging studies from 12 different centers (Ocaña-Tienda et al. (2023), Ramakrishnan et al. (2023), Wang et al. (2023)) to predict the origin of brain metastases from brain MRI scans. After using ensemble feature selection methods, we trained three fixed-effects and one mixed-effects multiclass classification model to account for longitudinal data. Fixed-effects logistic regression and support vector kernel models performed the worst (AUC 0.74 and 0.75, respectively), while the random forest performed better (AUC 0.97). The mixed-effects hierarchical Gaussian process boosting (GPBoost) model performed the best with an accuracy of 85.8%, precision of 85.4%, sensitivity of 85.8%, specificity of 95.7%, and an AUC of 0.98. In one-versus-rest classification, the GPBoost model identified the primary tumor as melanoma with the highest accuracy (96.1%) and non-small cell lung cancer with the lowest accuracy (89.0%). Developing such models holds great promise in reducing patient and institutional costs, improving time to effective treatment, and enhancing overall survival rates through personalized therapy approaches. In the future, we hope that incorporating newer patient data and refining the model to improve external validity would enable more accurate and generalizable predictions, leading to better clinical outcomes across diverse populations.