Beyond the naked eye: a systematic review on the current state of radiomics approaches to the vestibular schwannoma.

PURPOSE: Vestibular schwannomas (VS) present a clinical challenge in management decision-making due to their difficult-to-access location, unpredictable growth, and potential impact on crucial neurological function. This systematic review evaluates and summarizes the potential for radiomics, a computational tool that extracts quantitative features from imaging, to predict VS clinical outcomes and assess treatment responsiveness.

METHODS

Studies were extracted by searching PubMed, OVID Medline, and Web of Science databases. Included studies analyzed radiomic features from MRI as independent variables and varied in their methodology to predict clinical outcomes. Studies evaluated associations between radiomic features, pre-procedural clinical features, and post-procedural outcomes.

RESULTS

Thirteen retrospective studies met inclusion criteria; eleven of these used machine learning models to analyze radiomic MRI features. One non-ML study correlated longitudinal tumor volumetric changes with texture features. All segmentation workflows utilized manual or semi-automated approaches to determine the lesion's region of interest. Models based on pre-procedural imaging demonstrated moderate predictive accuracy by Area Under the Receiver Operating Characteristic curve (AUC = 0.66-0.7), while post-procedural models showed moderate to strong predictive capacity (AUC = 0.75-1.0). One study employed a convolutional neural network evaluating postoperative facial nerve outcomes (AUC = 0.89) that outperformed traditional ML models (AUC = 0.64-0.85).

CONCLUSION

Radiomics-based predictive modeling in VS shows encouraging preliminary results across a range of clinical outcomes. However, small sample sizes, retrospective designs, and lack of standardization and external validation in models hinder its widespread applicability. Addressing these limitations through prospective studies with standardized datasets and models, potentially incorporating deep learning, will be essential to improve generalizability and support clinical integration.