2026 Poster Presentations
P216: RADIOMICS AND MACHINE LEARNING FOR PREDICTION OF CAVERNOUS SINUS INVASION IN PITUITARY ADENOMAS
Maleeha Ahmad, MD, FRCS, FACS; UHS
Background: Cavernous sinus invasion (CSI) is one of the most critical determinants of surgical resectability in pituitary adenomas. Accurate preoperative identification of CSI informs surgical approach, extent of resection, prognosis and the need for adjuvant therapy. Conventional MRI features, such as Knosp grading, remain limited by sequence quality, interobserver variability and reduced sensitivity. Radiomics, which extracts quantitative features such as texture, shape, and intensity from MRI, combined with machine learning (ML) offers a reproducible and scalable method to improve CSI prediction.
Methods: A PRISMA-based review identified seven studies published between 2020 and 2025 that investigated radiomics and ML models for CSI prediction. Data were extracted on study design, model architecture, diagnostic accuracy and readiness for clinical integration.
Results: Radiomics-based classifiers consistently outperformed conventional MRI grading, achieving high AUC values of 0.80 to 0.92 in predicting CSI. Models that integrated clinical features with radiomics achieved the highest accuracy, consistently surpassing expert radiologist assessment alone. Radiomic signatures derived from tumor texture, boundary sharpness, and shape deformation correlated strongly with intraoperative findings of cavernous sinus invasion. Feature attribution techniques made the models interpretable by showing which MRI features, such as texture irregularity or boundary sharpness, contributed most strongly to a prediction of invasion. This makes the model clinically translatable by showing which MRI features drive the prediction of invasion, allowing surgeons to directly incorporate these insights into preoperative planning. Limitations across studies included retrospective design and the need for external validation across multiple centers.
Conclusion: Radiomics and ML represent a significant advance in the preoperative assessment of cavernous sinus invasion in pituitary adenomas. By providing automated, reproducible, and interpretable predictions, these models reduce diagnostic variability and enhance surgical planning. Integration into radiology workflows within 3 years is anticipated, thereby enabling surgeons to receive automated CSI predictions alongside MRI reports. This innovation has the potential to reduce morbidity, optimize extent of resection and refine preoperative strategy through precise risk stratification.
Keywords: Radiomics, Machine learning, Pituitary adenoma, Cavernous sinus invasion, Preoperative planning, Skull base surgery, MRI