2026 Proffered Presentations
S166: AN AI TOOL FOR ENHANCING SKULL BASE ANATOMY LEARNING THROUGH INTERACTIVE SPATIAL NAVIGATION EXERCISES: FEASIBILITY AND VALIDATION
Michael G Routhbaum, MD1; Antonio Bernardo, MD1; Alexander I Evins, MD, PhD2; 1Department of Neurosurgery, Oregon Health & Science University; 2Weill Cornell Medicine, Neurological Surgery
Introduction: Surgical neuroanatomy demands an integrated understanding of the complex spatial relationships between cortical, neural, vascular, bony, and dural structures. Traditional didactics and cadaveric dissection, while foundational, remain limited by access, repeatability, and cognitive load during real-time anatomical reasoning. We have previously demonstrated the pedagogical efficacy of 3D navigation-based fly-through instruction for teaching and learning complex topographic skull base anatomy, particularly in small-group settings guided by expert facilitators. The emergence of large language models (LLMs) introduces a novel, underexplored opportunity to extend individualized, structured neuroanatomical training beyond the classroom.
Objective: We evaluate the feasibility and utility of a custom programmed LLM for delivering guided, stepwise, navigation-based exercises and assessments to supplement surgical neuroanatomy instruction for individual students outside the classroom or lab setting.
Methods: We implemented a Generative Performance Trainer (GPT)-driven assessment platform, programmed with navigation-based querying principles, that allows users to navigate between distant anatomical points using contiguous structures of one or more specified tissue types. The model adheres to strict anatomical rules of contiguity and tissue-type constraints, designed to simulate 3D spatial navigation. Navigation exercises are generated adaptively with escalating complexity. Feedback and optional hints are provided in a tutoring-like manner. The exercises and answers are derived from a supplied knowledge base, including from lectures utilized in our neuroanatomy training program. Usability, internal consistency, anatomical validity, response grading accuracy, and the perceived educational value of the model were assessed.
Results: The Surgical Neuroanatomy Navigation Assessment Tool successfully generated anatomically valid, progressively complex questions across multiple cranial compartments and tissue types. Simulated use yielded >95% internal consistency and high anatomical accuracy in both query generation and answer validation across multiple sessions. The model reliably enforced contiguity and tissue-specific constraints and appropriately flagged incorrect user responses. The platform's adaptive difficulty and immediate feedback features enabled user-directed learning. No major anatomical inaccuracies or invalid question paths were identified across >150 queries. Minor inconsistencies were limited to terminology specificity and were correctable by backend rules refinement.
Conclusion: The integration of an interactive LLM-based navigation tool as an adjunct to skull base surgical anatomy learning is feasible and easily replicable. This tool enables high-fidelity, individualized practice of complex spatial navigation in skull base and neuroanatomy. Its low cost, ease of implementation, and content adaptability make it a compelling adjunct to traditional cadaveric and didactic instruction.