Open Pathology Learning Lab

A 2026 update anchored on two papers published this year in Nature Communications — EAGLE and a 32-model foundation-model benchmark — and an argument that both are excellent and conceptually incomplete. EAGLE’s lesson is orchestration, not two-thousandths of AUROC: a frozen model costing 0.36 ms decides where an expensive one looks. Thirty-two models later there is no winner, and that is the finding. Then the harder claim: the boundaries between asking a question, finding data, writing code, analyzing, publishing and teaching have collapsed — and what is scarce now is the question, the judgment, and the evidence. Closes on the Farmington Canal, whose right-of-way is still carrying traffic 200 years later.

Many labs can digitize slides; few can orchestrate. The image should serve the workflow, not define it. This 35-minute talk treats the WSI as a stateful production artifact and presents an event-driven, state-machine model that coordinates LIS events, the imaging lifecycle, distribution, and AI into one operational workflow.

The extended treatment for an academic informatics audience. Argues that “image management” is a category error, models the clinical case as a finite state machine with a nested per-slide FSM, and maps five decoupled layers — pixels, metadata, annotations, security, communication — onto open standards already in use. Draws the geospatial parallel (COG, Zarr, GeoJSON, OpenLayers) and works through Starling, Pelican, and Willet as one concrete instantiation.

AI-assisted development has changed what a small applied-informatics team can build inside an academic medical center. The talk traces the gap between operational EHR data and what clinical research actually needs, the parallel infrastructure departments quietly build to bridge it, and the institutional discipline that becomes both possible and necessary as iteration cycles collapse from months to hours. Includes live demonstrations of IBIS — natural-language search and cohort discovery — and Xenonym, a schema-driven synthetic-data generator. Closes with the case for applied informatics as the translational layer that turns biomedical information science into clinical reality.

The research-audience cut: from question to enterprise-grade tool in weeks. Opens with the cohort question that died in the analyst queue, then makes the case that what AI changed is not the quality bar but the cost of building. Distinguishes a vendor product, a service center, and an applied informatics team — and which is the right partner for which question — closing with a one-page specification you can draft for your own program.

Yale OneIT Conference — what changes, what doesn’t, and what it means for the act of building. Featuring the Xenonym case study, DevSecOps patterns, and the Tufte meta-argument.

How emerging technologies are transforming the tools we build — combining workflow foundations, the functionality gap, AI-assisted development evidence (Xenonym, IBIS, WBC ΔΣ), and a practical pathway for pathologists.

Logic, workflow, and the future of practice — featuring live polls, micro-case voting, D3 visualizations, and a JSON-driven slide engine.

Why we build, what we build, and how you can too — featuring the Buy vs. Build debate, problem-first clinical cases, hands-on workshop, live polls, and D3.js cost/architecture visualizations.

The original lecture — DevSecOps, image analysis tools, digital pathology infrastructure, and spotting workflow gaps.

What every pathology resident should know about LIS and IMS — covering HIPAA, regulatory frameworks, breach data visualization, and de-identification.