Computer-aided prediction of short-term tumor growth in sporadic vestibular schwannomas using both structural and dynamic-contrast enhanced MR imaging

Introduction:
Recent studies have demonstrated that microvascular parameters derived from dynamic-contrast enhanced (DCE) MR imaging significantly correlate with tumor growth in vestibular
schwannomas (VS). Other studies provide evidence that the use of artificial intelligence (AI)
on structural MR data provides similar predictive value for tumor growth.

Methods:
This prospective study investigates the combination of structural and DCE imaging data for AI to predict short-term tumor growth in VS. A total of 110 newly diagnosed unilateral sporadic VS patients underwent both T2-weighted and DCE MR imaging. Established pipelines were used to estimate the values of DCE-derived parameters Ktrans, ve, and vp. Subsequently, tumors were delineated and only voxel values within the delineation were considered for the AI model development. Radiomic features were extracted from both the structural images and DCE-derived parameter maps. A classifier was trained on the radiomic features to predict tumor growth.

Results:
Growth was observed in 69 (63%) of the 110 patients during follow-up. A support vector machine (SVM) model was trained on Ktrans and ve radiomic features using five-fold-crossvalidation. This model resulted in an accuracy of 82.5%, sensitivity of 81.2%, specificity of 82.9%, and area-under-the-curve of 0.85. The predictive value of structural MR imaging features is currently under investigation, as well as the use of more complex AI models. It is hypothesized that the addition of structural features and increase in model complexity will improve the model's predictive power. 

Conclusion:
Preliminary results have shown that DCE-derived parameter values exhibit a high predictive value for tumor growth prediction in sporadic VS. Other radiomic features and model types will be analyzed in order to investigate whether they improve the current AI model. These results will be presented during the conference.