This presentation by Dr. Matt Cecchini explores the transformative potential of artificial intelligence in pathology, framing it as a tool for augmentation rather than replacement. The core argument is that digitizing pathology slides is the essential first step to "unlocking the data behind the glass," making it accessible to powerful new AI tools. Dr. Cecchini introduces the concept of "agentic AI" and "reasoning models"—advanced AI systems capable of dynamic problem-solving, using tools, and performing complex tasks. The talk showcases a range of practical applications, from creating custom software with simple text commands to developing AI models on consumer-grade hardware for identifying cancer. Further examples include tools to simplify complex scoring like PDL1, using AI as a "thought partner" for differential diagnoses, and leveraging "stream of consciousness" dictation to auto-generate reports. The presentation culminates in a vision for a future where AI assistants seamlessly integrate into the pathology workflow, prioritizing cases, aiding in diagnosis, and automating tasks, thereby empowering a more efficient and capable future for the field.
Unlocking the Future of Pathology with Agentic AI
0:31 - Augmentation, Not Replacement: Dr. Matt Cecchini advocates for an AI-augmented future in pathology, emphasizing that professionals should focus on how they can be empowered by AI, rather than competing with it.
1:29 - AI Empowers Non-Coders: The rise of "vibe coding" tools allows users to generate functional applications, such as a PDL1 scoring training app, using simple text descriptions, democratizing software development.
3:31 - Digital Pathology as Foundation: The presenter stresses that merely switching from analog to digital microscopes is insufficient; true transformation comes from "unlocking the data behind the glass" through digital scanning, making it accessible for AI.
4:53 - Massive Data Scale: Pathology generates enormous amounts of data (e.g., 1.25 petabytes/year for one hospital), far exceeding other medical imaging modalities like radiology, presenting a vast, exciting resource for AI development.
6:43 - Agentic AI for Workflow Efficiency: "Agentic tools," defined as AI systems with dynamic control, information retrieval, tool usage, and memory, can parallelize and significantly streamline traditional, linear pathology workflows.
12:08 - Reasoning Models: A Breakthrough: Modern "reasoning models" represent a significant advancement, capable of "thinking" through problems step-by-step rather than just predicting, leading to dramatically improved accuracy and performance on complex tasks.
16:44 - Accessible AI for Clinical Problems: Practical AI solutions can be developed without high-end equipment; an example shows an effective basal cell carcinoma detection model trained on a consumer-grade gaming PC in weeks.
19:07 - Automating Report Generation: AI can be used to detect tumor margins and their relationship to ink, providing a foundation for "autocompleting pathology reports" by identifying tumor distribution, depth, and extent.
20:56 - Simplifying PDL1 Scoring: A specialized AI tool simplifies the challenging task of PDL1 reporting by automatically calculating the denominator (total tumor cells), making it easier for pathologists to determine the required positive cell count.
22:59 - AI as a Diagnostic Thought Partner: Reasoning models excel as "thought partners" for challenging cases, brainstorming differential diagnoses and suggesting IHC panels based on described scenarios, outperforming traditional search methods.
25:50 - "Stream of Consciousness" Dictation: Multimodal AI allows pathologists to dictate observations in a natural "stream of consciousness" while reviewing slides, which the AI then transforms into structured, synoptic reports, eliminating inefficient task-switching.
27:48 - Real-Time Interactive Assistant: A multimodal AI can observe the pathologist's screen in real-time, interpret images (e.g., gross descriptions, histopathology), and engage in interactive discussions, aiding in diagnosis and teaching.
30:50 - Insights from Eye-Tracking: Studies using eye-tracking reveal how pathologists visually analyze slides differently (e.g., faster, more logical scans for "Where's Waldo" or perceiving mitotic figures in peripheral vision), providing insights for better AI tool design.
33:44 - Creative Data Utility: Once digitized, pathology data can be repurposed creatively; an example demonstrates converting slide features into musical compositions, where untrained listeners could discern the transition from tumor to non-tumor.
35:08 - Integrated AI Vision for 2030: The future vision includes AI prioritizing cases, performing deep research, finding similar cases, and assisting with administrative tasks, seamlessly integrating into the workflow to create a more efficient and powerful pathology practice.