Computer-aided prediction of growth in vestibular schwannomas based on both structural and dynamic contrast-enhanced MR imaging

Background Vestibular schwannomas (VS) are benign intracranial tumors in the cerebellopontine angle. For most small-sized and medium-sized tumors, a wait-and-scan (W&S) approach is employed, since active treatment does not necessarily lead to symptom alleviation and up to 60% of VS tumors remain stable or regress during their natural course. At diagnosis, it is currently not possible to reliably predict tumor behavior and whether the tumor will grow or not, although some studies employing dynamic contrast-enhanced (DCE) magnetic resonance imaging (MRI) or artificial intelligence (AI) suggest potential evidence.Purpose Personalized clinical decision-making for VS patients may be improved if tumor growth can be predicted. This study therefore prospectively investigates the use of both structural T2-weighted MRI and DCE-derived microvascular biomarker maps (Ktrans, ve, and vp) in combination with machine learning techniques for the short-term prognosis of VS.Methods Newly diagnosed patients with unilateral sporadic VS and considered for an initial W&S strategy were enrolled between January 2021 and January 2023. Study participants underwent MR imaging, including both T2-weighted and DCE-imaging, within 3-6 months after diagnosis and were followed-up according to the usual standard-of-care regimen. The DCE-derived parameter maps (Ktrans, ve, and vp) were calculated using established pipelines. Radiomic features are extracted from the structural MR images and the three DCE-derived parameter maps. The resulting feature vector is reduced in dimensionality, using an F test and principal component analysis (PCA). A support vector machine (SVM) model is subsequently trained for the classification of tumor growth. The results of the F test are employed to assess the predictive value of each calculated feature.Results A total of 110 patients are analyzed for this study, of which 70 (64%) exhibited growth during follow-up. After fivefold cross-validation, the SVM model yields a mean accuracy of 89.0%, sensitivity of 90.0%, specificity of 87.7%, and AUC of 0.89 at the optimized cut-off probability for the prediction of tumor growth. The Ktrans and ve parameter maps provide most of the features to the model, comprising mostly of complex radiomics, as opposed to first-order statistics.Conclusions The developed model demonstrates the high potential of the combination of machine learning techniques and dynamic MRI for the prediction of tumor growth in VS patients. The DCE-derived parameters show high predictive value and provide insight into possible links between tumor biology and growth mechanisms. Complex radiomic features have revealed to be superior to first-order statistics as predictors for VS tumor growth, while preserving a degree of model explainability. Prior to clinical implementation, the reproducibility of the radiomics and model itself should be externally validated.