2026 Poster Presentations
P215: RADIOMICS AND MACHINE LEARNING FOR PREOPERATIVE PREDICTION OF PITUITARY MACROADENOMA CONSISTENCY
Maleeha Ahmad, MD, FRCS, FACS; UHS
Background: Pituitary macroadenoma consistency is a critical determinant of surgical complexity and extent of resection. As skull base surgeons we know, firmer tumors increase operative risk and likelihood of residual or recurrent disease, while softer tumors are more amenable to gross-total removal. Conventional MRI features are unreliable for predicting consistency. Radiomics extracts quantitative features (such as texture, intensity, heterogeneity) from MRI, while machine learning (ML) applies computational algorithms to these radiomics features to generate predictive models. Together, radiomics and ML offer the potential to provide non-invasive biomarkers of consistency to guide preoperative planning.
Methods: A PRISMA-based review identified five studies published between 2020 and 2025 evaluating radiomics and ML models for prediction of pituitary adenoma consistency. Extracted data included study design, diagnostic accuracy, interpretability, and readiness for clinical integration.
Results: A recent dual-center study developed a clinicoradiological-radiomics ML model achieving an AUC of 0.87, substantially outperforming conventional MRI assessment. Importantly, the model incorporated feature attribution methods, most notably SHAP (Shapley Additive Explanations). SHAP assigns a numerical score to each MRI feature, such as entropy, intensity, or heterogeneity, indicating how much it increased or decreased the likelihood of firmness. In practical terms, SHAP allows us surgeons to understand why the ML model predicts “firm” or “soft”.
Four additional single-center studies confirmed that key MRI-derived features, such as entropy, gray-level co-occurrence, and volumetric heterogeneity, correlate with intraoperative firmness, though sample sizes were smaller and designs retrospective. Across all five studies, ML classifiers consistently outperformed expert radiologists using current MR protocols.
Conclusion: Radiomics and ML provide reliable, non-invasive prediction of pituitary macroadenoma consistency, directly informing surgical strategy and patient counseling. Interpretability methods such as SHAP are essential for clinical adoption as they explain model reasoning in the same way surgeons approach differential diagnosis where identifying which imaging features drives the prediction. This transparency builds surgeon confidence, aligns AI with preoperative planning, and strengthens its role as a decision-support tool. Integration at the MRI workstation level within the next 3 years is anticipated, where automated tumor segmentation and consistency prediction would be regularly incorporated into radiology reports. This translation has the potential to reduce surgical morbidity, enable tailored resection strategies, and ultimately serve as a bridge toward AI-robotic integration in skull base surgery.
Keywords: Radiomics, Machine learning, Pituitary macroadenoma, Consistency, Preoperative planning, Skull base surgery, MRI