Abstract 9
Category: Clinical Science
At the end of the session, participants will be able to:
- Explore how digital classification tools may support rapid intraoperative decision-making in CNS tumor surgery.
- Understand the role of patch-based image clustering in improving frozen section interpretation.
- Review a pilot framework for developing and validating a frozen section model across multiple CNS diagnostic categories.
COI Disclosure:
None to disclose.
Presenter
Adrienn Bourkas is a PGY1 neuropathology resident at the University of Toronto. Born in Budapest and raised in Toronto, she completed her BSc and MSc in Biochemistry, with research experience in neurodegeneration and rare hematologic disorders at the Tanz Centre and SickKids. She earned her MD at Queen’s University, where she was involved in projects ranging from AI diagnostics to oncology and orphan diseases. Her current interests include neuro-oncology, molecular diagnostics, and the future of personalized pathology.
Adrienn N. Bourkas1, Shane Eaton1, Phedias Diamandis1-4
1 Department of Laboratory Medicine and Pathobiology, University of Toronto, Toronto, ON, Canada
2 Princess Margaret Cancer Centre, Toronto, ON, Canada
3 Department of Medical Biophysics, University of Toronto, Toronto, ON, Canada
4 Laboratory Medicine Program, Department of Pathology, University Health Network, Toronto, ON, Canada
Target Audience:
Pathologists, Residents
CanMEDS:
Medical Expert (the integrating role)
Validation framework for an AI decision support tool in intraoperative neuropathology: Protocol for a multi-phase evaluation
Abstract
Intraoperative consultations (IOC) for neurosurgical cases rely on rapid assessment of H&E-stained frozen sections and cytologic preparations, often under time pressure and with variable tissue quality. We describe a validation framework for evaluating OnSightPathology, an open and platform-agnostic artificial intelligence (AI) tool developed to classify intraoperative digital slides into five diagnostic categories: glioma, epithelioid, meningeal, schwannoma, and normal/reactive tissue.
In the pilot phase, 10-20 representative cases from each class were selected from archival material, and 50-100 diagnostic tiles per case were extracted using an image feature based clustering pipeline. Groups of tiles were assigned coarse labels (e.g., glioma, artifact) via visual histologic inspection. Ground-truth diagnoses were adjudicated on formalin-fixed paraffin-embedded sections through dual-review consensus. The annotated images will then be used to train a vision transformer model and will be deployed using OnSightPathology to provide “real-time” inferences to help with intra-operative frozen section analysis.
The tool is designed to assist pathologists by highlighting diagnostically relevant regions, providing class probabilities, and flagging low-confidence areas for review. This approach is intended to augment IOC workflows by improving confidence, reducing turnaround time, and standardizing interpretation in time-sensitive settings. Subsequent phases will assess OnSightPathology’s performance on external cohorts and in real-time clinical use.