Dr. Matt Cecchini's presentation, "Unlocking the data behind the glass to empower the agentic future of pathology," advocates for digital pathology as the foundational step for integrating advanced AI tools into clinical practice. He emphasizes an "AI-augmented" rather than "AI-replacement" future, showcasing how large language models (LLMs) and agentic AI can transform pathology workflows. The talk delves into the capabilities of current AI for code generation, computer interaction, and sophisticated reasoning, while acknowledging their limitations. Dr. Cecchini presents practical applications, including automated tumor detection (e.g., basal cell carcinoma on frozen sections), margin analysis, and tools for specific scoring challenges like PDL1. He highlights AI's role as a "thought partner" for challenging cases and a deep research assistant. The presentation also explores multimodal AI for real-time dictation and image analysis, as well as novel applications like eye-tracking studies on pathologist behavior and even converting pathology slides into music, ultimately arguing that digitization of data enables a far more efficient, insightful, and "agentic" pathology workflow for the future.
Exploring the Agentic Future of Pathology with Digital Data and AI
0:00 Vision for Pathology's Future: Dr. Matt Cecchini introduces the concept of unlocking data in pathology to empower an "agentic future," emphasizing AI as an augmentation, not a replacement, for human pathologists.
1:29 AI's Practical Capabilities: Showcases current AI tools, including "vibe coding" for app development (e.g., a PDL1 scoring training app built with Replet) and AI's ability to interact with computers for tasks like scheduling.
2:52 AI Limitations: Highlights that while powerful, AI agents can spectacularly fail, citing an example where an AI task veered off course to randomly search for Yellowstone National Park.
3:30 Digital Pathology as Foundation: Argues that simply viewing slides on a screen isn't transformative; digital scanning is the essential first step for digital transformation, unlocking data for AI applications.
4:53 Immense Data Volume: Pathology generates massive amounts of data (e.g., 1.25 petabytes/year for one hospital, 10x Netflix library size), which, while daunting for administrators, is exciting for AI integration.
6:01 Pathology as a "Factory" Augmented by AI: Conceptualizes pathology as a factory (specimens to reports) and proposes "agentic tools" to parallelize and streamline the traditionally linear workflow.
7:08 Understanding AI Agents: Defines AI agents as systems where Large Language Models (LLMs) dynamically direct their processes, retrieve information, use tools, and learn from memory, noting their exponential improvement in task complexity.
9:01 AI Terminology Explained: Clarifies AI hierarchy (AI > Machine Learning > Neural Networks > Deep Learning) and explains LLMs (Generative Pre-trained Transformers) by analogy to pathology workflow, emphasizing their ability to connect distant information.
11:43 The Power of Reasoning Models: Illustrates how "reasoning models" overcome basic LLM failures (e.g., counting 'R's in 'strawberry') by spelling out problems and "thinking" longer, leading to vastly improved performance.
14:18 AI in Pathology - Early Successes: Recounts an early success where an AI (GPT-4) correctly identified lung tissue and H&E stain, exceeding expectations for a senior pathology resident.
15:11 Guiding Principles for AI Use: Advises caution, reminding users that LLMs can "hallucinate" and provide confident but wrong answers, stressing user responsibility, avoiding patient data (unless institutionally approved), and the rapid pace of AI improvement.
16:02 Practical AI Deployment Example (BCC): Demonstrates building an AI tool on a consumer gaming PC using an open-source model (Verkow) to accurately detect basal cell carcinoma on frozen sections from just 30 training cases, potentially aiding surgeons.
19:13 Agentic Workflows & Reporting: Describes an algorithm for margin detection (tumor touching ink) and envisions AI auto-completing pathology reports by identifying tumor, its relationship to margins, and measuring depth/extent.
19:53 Streamlined Analysis with Biomarker Analysts: Discusses using simpler image analysis tools for Ki67, HER2, PDL1, and MMR, supported by dedicated "biomarker analysts" (MLTs with AI training) who push results to pathologists.
21:09 PDL1 Reporting Tool: Presents a specific tool to simplify PDL1 scoring by calculating the denominator, aiming to reduce inter-observer variability among pathologists.
22:50 AI as a Pathologist's Thought Partner: Describes using reasoning models for brainstorming challenging cases (e.g., high-grade tumor with random stains, SmartA4 deficient tumor) and for efficient deep research, citing academic papers.
26:01 Multimodal AI for Workflow Transformation: Introduces "stream of consciousness dictation" (AI converting spoken thoughts into structured pathology reports) and real-time multimodal assistants (e.g., Gemini analyzing screen content to identify cancer types from histology).
29:57 Fun & Insightful AI Applications: Showcases creative AI uses like remixing images (UIP as letters, bee in honeycomb) and converting pathology slides into songs, demonstrating the versatility of digitized data.
30:28 Understanding Pathologist Behavior: Highlights eye-tracking studies revealing how pathologists search (faster, more logical scans than non-pathologists) and how they detect mitotic figures (using peripheral vision).
32:32 Screen Size Impact: Explores a study showing that while larger monitors (e.g., 24-32 inches) perform well, pathologists' eyes may not fixate on the outermost areas, suggesting that "bigger is not always better."
35:07 The Agentic Future of Pathology: Concludes with a vision of AI prioritizing cases, conducting deep research, finding similar cases, and assisting with administrative tasks, emphasizing digital pathology as essential for the field's advancement.