Beyond the Scope Dr. Faisal Mahmood, Harvard Medical School
Feb 2, 2026
Dr. Faisal Mahmood, Associate Professor of Pathology at Harvard Medical School and leader in computational pathology, discusses foundational AI models like UNI and CONCH. He covers slide- and patient-level representations, multimodal contrastive approaches, generative AI for morphology and triage, and agentic tools such as PathChat for interactive slide workflows.
AI Snips
Chapters
Transcript
Episode notes
Feature Extraction Drives Pathology Models
- Faisal Mahmood found feature extraction is the critical component in multiple-instance learning for pathology models.
- His lab moved from patch-level models to slide-level vectors to enable retrieval and downstream tasks.
Image And Contrastive Foundation Models
- UNI is an image-only foundation model for region feature extraction and CONCH contrasts image and text to improve representation.
- Slide-level models produce single vectors usable for search, supervised tasks, and clinical applications.
Power Of Multimodal Contrastive Training
- Multimodal contrastive training uses auxiliary views like reports or molecular data to enrich pathology representations.
- Molecular data provides a more objective second view than text reports for improving image features.