Abstract 8
Category: Basic Science
At the end of the session, participants will be able to:
- Introduce the applications and potential benefits of artificial intelligence in pathology for improving
diagnostic accuracy and workflow efficiency. - Evaluate the performance of AI tools in tumor classification, mitosis detection, and Ki-67 quantification,
and its potential impact on diagnostic accuracy and workflow efficiency.
COI Disclosure:
None to disclose.
Presenter
Parsa Babaei Zadeh is a second-year MD student at the University of Alberta with a growing focus on the intersection of artificial intelligence and pathology. He is especially interested in how machine learning can support diagnostic accuracy, accelerate digital pathology workflows, and expand access to high-quality care. Before medical school, Parsa completed his undergraduate studies at the University of Toronto, majoring in Biology and Immunology, where he built a strong foundation in cellular mechanisms, host–pathogen interactions, and experimental design.
Kevin Faust1, Jinzhen Hu1, Parsa Babaei Zadeh1, Adrienn Bourkas1, Dimitrios Oreopolous2, Phedias
Diamandis1,2
1. Princess Margaret Cancer Centre, 101 College Street, Toronto, ON, Canada.
2. Department of Laboratory Medicine and Pathobiology, University of Toronto, Toronto, ON, Canada.
Target Audience:
Pathologists, Residents, Medical Students
CanMEDS:
Medical Expert (the integrating role), Leader, Health Advocate, Scholar, Professional
OnSight: A real-time computational pathology companion for histopathology
Abstract
Objective: Precise microscopic examination of surgical tissue sections is a specialized skill critical for diagnosis and clinical care. While artificial intelligence (AI) shows promise for histological analysis, differences in slide digitization and proprietary software limit real-world deployment. We present OnSight, a platform-agnostic computer vision software providing real-time AI inferences to pathologists during digital slide review. OnSight is freely accessible as a single executable file, runs locally on consumer-grade computers, and requires no complex integration, enabling cost-effective and safe use in research and clinical workflows.
Methods: OnSight was benchmarked on classification, mitosis detection, and immunohistochemical quantification. For tumor classification (glioma, meningioma, schwannoma, epithelial metastasis), a ViT-B/16 pretrained on the Kaiko foundation model was fine-tuned on 80,000 H&E tiles from UHN at 20×. Ki-67 nuclei were segmented with a YOLO model trained on QuPath-annotated patches from 10 cases. Mitotic figures were detected with a RetinaNet FPN (ResNet-50 backbone) trained on MIDOG++ at 40×. Validation used TCGA and UHN cases.
Results: Tumor classification accuracy was 95% across public and institutional sets. Mitosis detection showed high area under the curve and agreement with manual review. Automated Ki-67 indexing showed closer agreement with neuropathologist counts than QuPath and is reported as a percentage. Median per-field latency was 0.3 s, enabling real-time use.
Conclusions: OnSight provides accurate, SI-unit-standardized, low-latency pathology slide analysis with potential to reduce inter-observer variability, improve accuracy, and increase efficiency. Prospective multicenter evaluation is warranted for widespread use.