P13.06.A RADIOMIC FEATURES FOR RISK-STRATIFICATION IN PATIENTS WITH BRAIN METASTASES

• Published in: Neuro-oncology (Charlottesville, Va.) Vol. 25; no. Supplement_2; pp. ii101 - ii102
• Main Authors: Heugenhauser, J, Herbst, S, Drucks, T, Scherfler, C, Iglseder, S, Urbanic Purkart, T, Hassler, E M, Kerschbaumer, J, Freyschlag, C, Wagner, M, Gizewski, E R, Nevinny-Stickel, M, Stockhammer, G, Nowosielski, M
• Format: Journal Article
• Language: English
• Published: US Oxford University Press 08-09-2023
• Subjects: POSTER PRESENTATIONS
• ISSN: 1522-8517; 1523-5866
• DOI: 10.1093/neuonc/noad137.340

Abstract
BACKGROUND
Recently, non-invasive characterization of brain tumors on MRI has emerged as a promising field of research. The identification of quantitative imaging biomarkers, also known as radiomics, may complement molecular characterization and thereby improve clinical management of neuro-oncologic patients. In this study we aimed to identify imaging predictors with improved performance over clinical parameters in order to stratify patients with brain metastases into high and low risk groups for overall survival (OS).
MATERIAL AND METHODS
422 patients (allocated in a 3:1 ratio to a discovery [n=317] and test [n= 105] set) with first diagnosis of brain metastases from different primary tumors from two neuro-oncologic centers were included. In each patient, eight clinical features (age, gender, KPS, systemic disease status, presence of extracranial metastases, number of cerebral metastases, primary tumor and available individual prognostic molecular status eg HER2, BRAF,⋯) were gathered and a total of 321 radiomic MRI features (including shape, first-order and higher-order features) from cerebral MRI (contrast T1-weighted and apparent diffusion coefficient maps) were extracted. Radiomic and clinical features of patients in the discovery set were subjected to different machine learning models in order to classify patients into low- and high-risk groups for OS. By performing a comprehensive grid search the best machine learning model according to the macro-evaluated score and accuracy was identified and evaluated on the testing set. In addition, a subgroup analysis, based on the primary tumor entity was done and confusion matrices were calculated in order to evaluate final predictions.
RESULTS
With an extra trees classifier including all clinical and 30 radiomic features, we were able to stratify patients into a high- and low-risk group for OS (test set: macro-evaluated = 0.60, accuracy = 0.67). The best performing model was a gradient boosting model only including clinical features (test set: macro-evaluated = 0.62, accuracy = 0.72), while the radiomic features alone led to the poorest results (test set: macro-evaluated = 0.52, accuracy = 0.63). Interestingly, in the subgroup of melanoma patients, the predictive power of radiomic features outperformed the clinical and the combined (radiomics and clinical) features. With a prediction solely based on radiomic features, 80% and 67% of patients were correctly classified into the low- and the high-risk group for OS, respectively.
CONCLUSION
In conclusion, we found that in the entire study population of patients with brain metastases radiomic features did not allow for a better prediction of the clinical outcome compared to the clinical parameters alone. However, in the subgroup of melanoma patients the predictive power of radiomic features alone was superior compared to clinical features or all features combined.