This presentation explores the transformative potential of Artificial Intelligence (AI) in digital pathology, advocating for an "AI-augmented" rather than "AI-replacement" future. The speaker, Matini, demonstrates the capabilities of current AI tools, including their ability to generate code (e.g., a PDL1 scoring app), interact with computers, and perform complex reasoning tasks. He emphasizes that digital scanning is merely the first step in unlocking vast amounts of pathological data, which far exceeds that of other medical imaging fields. Matini showcases practical applications, such as using foundation models for rapid basil cell carcinoma detection, automating report generation, and assisting with complex diagnoses and research queries through "deep research" AI. He also highlights multimodal AI capabilities, like converting stream-of-consciousness dictation into structured reports and real-time interactive case review. The talk touches on research into pathologist eye-tracking and even creative data uses like converting pathology slides into music. The overarching vision for 2030 is a pathology workflow where AI prioritizes cases, acts as a thought partner, and assists with various administrative and research tasks, underlining the critical need for digital pathology infrastructure to leverage these advancements.
AI in Pathology: Unlocking the Data Behind the Glass
0:33 AI Augmentation, Not Replacement: The speaker believes in an AI-augmented future where AI tools enhance human capabilities, rather than replacing jobs.
1:31 AI for Code Generation: Demonstrates using a "no-code" tool (Repl.it) to build a functional app for training PDL1 scoring with simple text prompts.
2:30 AI Utilizing Computers: Shows AI agents performing internet searches and calendar integration, indicating potential for workflow transformation.
2:55 AI Limitations and Hallucinations: Cautions that while AI can be powerful, it can also spectacularly fail and provide confidently wrong answers (e.g., Yellowstone National Park example).
4:08 Digital Scanning as the First Step: Emphasizes that digital scanning is crucial for unlocking the data needed for AI tools, as analog microscopes offer "zero compatibility."
5:03 Vast Pathology Data: Pathology data volume is immense, projected to be 10 times the size of Netflix's library annually for one hospital and potentially exceeding Google Maps data within 20 years.
6:01 Pathology as a Factory, Augmented by Agents: Conceptualizes pathology as a factory where AI agents can streamline processes, moving cases through the pipeline more efficiently.
7:10 Understanding AI Agents: Defines AI agents as systems using large language models (LLMs) to dynamically direct their own processes, retrieve information, use tools, and learn from memory.
8:02 Exponential Growth of Agentic AI: Highlights the rapid improvement in AI's ability to complete complex tasks, now capable of tasks that would take a human 1-2 hours.
9:06 AI Terminology Simplified: Explains AI, Machine Learning, Neural Networks, and Deep Learning as a nested hierarchy, with LLMs being generative pre-trained transformers.
11:09 Tokenization and Long-Range Connections: Explains how LLMs break down language into "tokens" and can connect them over very long ranges, enabling powerful context understanding.
12:08 Reasoning Models - A Key Shift: Explains that modern "reasoning models" don't just "spit out an answer" but actively "think about the problem," leading to significantly better performance (e.g., correctly counting "R"s in "strawberry").
13:10 Overview of Current LLMs: Briefly lists major LLM models from OpenAI (GPT series), Google (Gemini), Anthropic (Claude), X (Grock), Meta (Llama), and Deepseek.
14:24 AI in Pathology - Early Successes: Recounts giving an AI a histology image and it correctly identified lung tissue, H&E stain, and noted it was normal.
15:13 Guiding Principles for AI Use: Stresses caution against hallucinations, user responsibility, avoiding patient data (unless compliant), and the rapid pace of AI improvement.
16:10 Practical Application - Foundation Models (Verkow): Details a study using the open-source Verkow model on a low-cost gaming PC to identify basal cell carcinoma (BCC) on frozen sections, correctly mapping tumor areas.
19:12 Practical Application - Workflow Automation: Demonstrates an algorithm detecting tumor margins and localizing tumor, envisioning auto-completion of pathology reports.
19:53 Practical Application - Simpler Image Analysis: Discusses in-house tools for P67, HER2, PDL1, and molecular testing, supported by a "biomarker analyst" role.
21:11 Practical Application - PDL1 Scoring Aid: Presents a tool calculating the denominator for PDL1 scoring, aiming to improve inter-observer variability.
22:32 Practical Application - AI as a Thought Partner: Describes using reasoning models to brainstorm differentials for challenging cases or help decide on IHC panels without inputting patient data.
23:56 Practical Application - Deep Research Tools: Highlights AI with search and reasoning capabilities for efficient literature review and preparing talks on unfamiliar topics, citing sources.
25:08 Practical Application - Perplexity Examples: Gives real-world examples of using AI search (Perplexity) to quickly find clinical information (e.g., MDM2 FISH testing criteria for lipoma, sebaceous carcinoma insights).
26:06 Practical Application - Multimodal AI (Stream of Consciousness Dictation): Explores converting natural, unstructured dictation during microscope review into a structured pathology report, significantly improving efficiency.
27:48 Practical Application - Multimodal Assistant Demo: Showcases Google's Gemini AI interacting in real-time by viewing the speaker's screen, understanding pathology reports, and identifying cancer types from images.
30:31 Research - Finding Rare Elements (Where's Waldo): Discusses AI's ability to find rare tumor fragments and lab research using eye-tracking to understand how pathologists find specific elements.
31:09 Pathologist Eye Gaze vs. Non-Pathologist: Reveals that pathologists scan more logically and fixate for shorter durations, making them better "Waldo finders."
31:47 Peripheral Vision for Motic Figures: Pathologists can detect motic figures in their peripheral vision, unlike "Where's Waldo," indicating different cognitive processing for specific tasks.
32:34 Research - Optimal Screen Size: A study suggests that excessively large monitors might not be better for digital pathology, as pathologists' eyes may not fully engage the extreme edges.
33:38 Creative Data Use - Pathology Songs: Presents a unique project where pathology slides are converted into musical compositions, demonstrating the versatility of digitized data.
35:09 Vision for 2030 Pathology: Envisions AI prioritizing cases, acting as a thought partner, searching archives for similar cases, and assisting with research tasks like writing REBs.
35:41 Call to Action: Encourages adoption of digital pathology to keep pace with the rapidly advancing AI landscape, emphasizing that many AI tools are already accessible.