Episodes

  • 249: Cytopathology AI: Crowded-Cell Gaps and LLM Guardrails

    249: Cytopathology AI: Crowded-Cell Gaps and LLM Guardrails
    Aug 31 2026
    Send us Fan MailWhat happens when AI models that perform almost perfectly on scattered cervical cells become less reliable than a coin flip on crowded cell groups?In DigiPath Digest #51, I examine what this performance gap tells us about artificial intelligence in cytopathology.The first paper evaluated six convolutional neural network models trained to distinguish benign from high-grade lesions using scattered cervical cytology cells. The models achieved AUCs ranging from 0.950 to 0.996 on the original images.When the same models were applied to hyperchromatic crowded cell groups without retraining or threshold recalibration, their AUCs fell to 0.385-0.683. Different architectures also failed differently. Some overcalled benign clusters, while others missed high-grade clusters.This isn’t simply a technical problem. It illustrates a practical rule for pathology AI: a model should only be trusted for the morphology, specimen type, imaging system, and intended use on which it has been directly validated.I also review the current evidence for large language models in cytopathology. Potential applications include structured reporting, diagnostic support, quality control, education, research, and workflow integration.Some early results appear promising, but each comes with important limitations. One structured-reporting application achieved 99.4% accuracy at a single institution. A diagnostic-support model included the correct answer among its top 10 differentials in 59.1% of general medicine cases. A hybrid quality-control system flagged 84% of errors associated with amended reports, but its false-positive rate wasn’t reported.Most importantly, the review found no language model specifically trained and clinically validated on cytopathology reports.The takeaway is straightforward: we’re still working with narrow AI. Strong performance in one setting doesn’t guarantee performance when the cells, preparation, scanner, institution, or clinical task changes.Low-risk applications may offer the most practical starting point. Text extraction, completeness checks, report consistency review, and quality-control flagging could reduce repetitive work without asking an unvalidated model to make the final diagnosis.Highlights with timestamps00:00 - Welcome to DigiPath Digest #51 and the new lunch-and-learn time01:40 - How two image models performed worse than a coin flip on cell clusters02:29 - Image models, language models, and vision-language models04:40 - Why AI adoption in cytopathology remains low05:25 - Scattered single cells versus hyperchromatic crowded cell groups08:14 - AUCs fall from 0.950-0.996 to 0.385-0.68310:02 - How ResNet-50 and GoogLeNet failed differently11:04 - What the attention maps revealed13:22 - The intended-use lesson for pathology AI17:02 - Current applications of large language models in cytopathology18:39 - Structured reporting and the 99.4% accuracy result19:32 - Diagnostic support, AMIE, and the top-10 limitation21:09 - Quality-control applications and the missing false-positive rate23:30 - Why cytopathology still needs domain-specific language models24:54 - Retrieval-augmented generation, education, and research support27:07 - Two AI families, one requirement: direct validation28:35 - Context of use and intended-use validation29:12 - Human-reviewed training data and destructive book scanning32:02 - FDA and European approaches to evolving AI models35:00 - Protecting patient and practitioner well-being36:21 - Cytopathology-specific benchmarks and shared test sets38:16 - Why low-risk AI applications should come first39:50 - Cytopathology’s direct-to-digital advantage40:39 - Digital Pathology 101 and Pathology VisionsResources from this episodeWatch DigiPath Digest #51Diagnostic performance of AI models trained on scattered single-cell imagesLeveraging large language models to enhance cytopathologyFDA guidance on AI and context of useFDA guidance on predetermined change control plansGoogle Research: AMIE diagnostic medical AIEU In Vitro Diagnostic Medical Devices RegulationEU Artificial Intelligence ActDigital Pathology 101Pathology Visions 2026 Support the showGet the "Digital Pathology 101" FREE E-book and join us!
    Show More Show Less
    31 mins
  • 248: Are Foundation Models Really Better for Digital Pathology? Podcast with Panu Kauppila

    248: Are Foundation Models Really Better for Digital Pathology? Podcast with Panu Kauppila
    Aug 20 2026
    Send us Fan MailWhat good is a powerful foundation model if it slows the pathologist down, can’t explain its result, or doesn’t fit the clinical workflow?Foundation models are gaining attention across digital pathology. But they’re not finished clinical tools by themselves.In this episode of the Digital Pathology Podcast, I speak with Panu Kauppila, Chief Product Officer at Aiforia, about what foundation models are, how they differ from traditional convolutional neural networks, and what it takes to make them useful for pathologists.Panu describes a foundation model as a large, context-aware building block. To perform a specific pathology task - such as grading, segmentation, or mitotic counting - it must be combined with an adapter, a task-specific head, curated annotations, a usable interface, and integration with the laboratory workflow.We also discuss one of the biggest practical constraints: speed.A pathologist shouldn’t have to click a button and wait for an analysis. Panu explains why AI should run automatically in the background so the results are already available when the case reaches the pathologist’s worklist.The conversation also examines the tradeoff between model size, computational cost, and clinical performance. Larger models may improve robustness and generalizability, but they can also require more processing power. For clinical applications, Aiforia focuses on smaller and medium-sized foundation models that provide the necessary quality without making the workflow slower or unnecessarily expensive.Annotated data remains central. The foundation model supplies the underlying image understanding, while the task-specific head and controlled annotations determine how the model performs on a particular pathology problem. This structure also raises important questions about bias, data provenance, ownership, regulatory documentation, and explainability.Finally, we look at multimodal AI models that combine pathology images with reports, genomic data, molecular information, and clinical outcomes. These tools could support more interactive, predictive, and prognostic applications - but only if they’re introduced through secure, controlled workflows with clear audit trails.Episode Highlights00:00 — Where does bias enter a foundation model workflow?Panu distinguishes the underlying foundation model from the annotated dataset used to build the task-specific application.00:27 — Meet Panu KauppilaAn introduction to Aiforia’s Chief Product Officer and the episode’s focus on foundation models in digital pathology.01:05 — From radiology AI to digital pathologyPanu describes his background in medical device development, radiology, oncology solutions, and clinical AI implementation.06:29 — Foundation models versus convolutional neural networksWhat makes foundation models more context-aware, robust, and generalizable across image datasets.07:07 — Image-only and multimodal foundation modelsWhy these two categories offer different capabilities and potential clinical uses.07:50 — A foundation model is a platform, not a finished solutionThe underlying model may understand image features, but it still needs a task, interface, and clinical workflow.08:34 — Foundation model, adapter, and task-specific headHow these components work together to create an application for grading, mitotic counting, or another pathology task.09:57 — The cost of larger modelsWhy increased robustness must be balanced against computational demands, inference speed, and affordability.11:12 — Pathologists won’t wait for AIWhy even short delays can interrupt the clinical workflow.11:46 — Running AI in the backgroundA workflow in which slides are scanned, analyzed automatically, and added to the worklist with results ready for review.12:16 — Combining foundation models with curated annotationsHow smaller task-specific datasets and adapter technology can produce practical pathology models.15:24 — Generalizability across scanners, laboratories, and populationsHow foundation models may make adaptation to new domains more manageable.16:36 — Two datasets, two sources of potential biasThe regulatory questions created by an underlying foundation model and a separate controlled annotated dataset.20:46 — Making foundation models accessibleWhy a platform and user interface are necessary for pathologists and researchers who don’t work directly with code.21:57 — Testing foundation models in Aiforia CreateHow researchers can compare a CNN with supported foundation models in the same no-code environment.25:45 — How foundation models are selectedQuality, licensing, model size, annotated data, and the requirements of the intended use.26:57 — Training cost versus inference costWhy a more expensive training iteration may still reduce total development costs if fewer iterations are needed.32:23 — Pathologists are visual reviewersThe importance of segmentation quality and showing exactly what the model ...
    Show More Show Less
    53 mins
  • 247: Screening Efficiency Over Experience: Rethinking Cytology Expertise

    247: Screening Efficiency Over Experience: Rethinking Cytology Expertise
    Aug 17 2026

    Send us Fan Mail

    Does more experience automatically make a cytotechnologist more accurate—or does where they look first matter more?

    In DigiPath Digest #50, I review a digital cytology eye-tracking study that challenges the assumption that diagnostic accuracy improves steadily with years of practice.

    The researchers tracked the visual behavior of 100 board-certified cytotechnologists with 1 to 40 years of experience. They found no statistically significant linear relationship between years of experience and diagnostic accuracy. Instead, low-power field efficiency—the ability to identify an important target quickly within a wider field—emerged as the key predictor of high accuracy discussed in the study.

    The study also examined whether this visual skill can be developed. Twenty-eight students completed an intensive three-month cytotechnology training program. After training, they located diagnostic targets more quickly and spent less attention on normal, nondiagnostic cells. In other words, they learned both where to look and what to disregard.

    What could this mean for digital pathology education?

    As AI-assisted workflows take on more of the exhaustive searching, cytotechnologists and pathologists may increasingly work as expert verifiers. That requires rapid target assessment, strong knowledge of normal morphology, and awareness of risks such as confirmation bias and cognitive fatigue.

    The study has an important limitation: it used static images rather than dynamic whole slide imaging. The findings raise useful questions about visual expertise, training, and competency assessment, but they shouldn’t be generalized beyond the study design without further research.

    Episode Highlights

    • 00:00 – Welcome to DigiPath Digest #50 and introduction to the paper
    • 04:10 – Why the traditional definition of professional expertise is changing
    • 07:02 – Moving from exhaustive searching to verification in AI-assisted workflows
    • 09:04 – How eye tracking was used with 100 board-certified professionals
    • 10:25 – Years of experience versus diagnostic accuracy
    • 13:13 – Experience-based caution and attention to sample information
    • 15:16 – Low-power field efficiency as a predictor of high accuracy
    • 17:05 – Practical low-power field demonstration using a whole slide image
    • 20:15 – Searching versus detecting and the mental map of normal morphology
    • 24:04 – Comparing high- and low-performer visual scan paths
    • 25:27 – Cognitive filtering: knowing what not to examine
    • 27:35 – Can visual efficiency be taught in three months?
    • 29:55 – How AI may shift the human role from searcher to verifier
    • 30:38 – Study limitations: static images versus dynamic whole slide imaging
    • 32:37 – Could gaze efficiency influence future competency assessment?
    • 33:44 – Digital pathology learning resources and closing thoughts

    Resources Mentioned

    • Abstract and paper: Screening Efficiency Over Experience: Rapid Target Detection in Low-Power Field as a Modifiable Cognitive Biomarker for Diagnostic Accuracy in Digital Cytology

    Listen to the full DigiPath Digest #50 recording to examine what the study found, what it didn’t prove, and how visual search skills could influence digital cytology training.

    Support the show

    Get the "Digital Pathology 101" FREE E-book and join us!
    Show More Show Less
    29 mins
  • 246: Computational Pathology Is Changing Companion Diagnostics

    246: Computational Pathology Is Changing Companion Diagnostics
    Aug 13 2026
    Send us Fan MailCan a treatment decision depend on whether one pathologist sees 45% biomarker positivity and another sees 55%?Visual immunohistochemistry scoring helped establish precision oncology. But as targeted therapies become more sensitive to subtle biological differences, categorical scores such as 0, 1+, 2+, and 3+ may no longer capture the information needed to identify the right patients.In this episode, I speak with three Roche experts:Gordana Juric-Sekhar, MD, anatomic pathologistSaleh Miri, PhD, Director of Digital Pathology AI AlgorithmsPurvi Gaglani, Regulatory Affairs Lead for Digital PathologyWe discuss how computational pathology is changing companion diagnostics by moving biomarker assessment from visual estimates to continuous, cell-level measurements.The conversation examines the limitations of manual IHC scoring, including interobserver variability, intraobserver variability, visual fatigue, borderline cases, tumor heterogeneity, and the inability of the human eye to measure complex spatial relationships.Using TROP2 scoring in advanced non-small cell lung cancer as an example, Saleh explains the normalized membrane ratio. This computational metric measures protein expression at the cell membrane relative to total expression within the cell—something that can’t be reproduced through conventional visual scoring.We also clarify the difference between computer-assisted scoring and a fully computational companion diagnostic. An assisted tool supports a pathologist’s visual interpretation. A computational CDx generates the biomarker measurement through algorithmic, cell-level analysis.That doesn’t remove the pathologist.Pathologists remain responsible for evaluating tissue quality, staining quality, scan quality, tumor selection, image analysis results, and the final clinical context. They can reject a stain, request a rescan, exclude inappropriate regions, question the result, or seek a second opinion.The episode also examines the regulatory implications of computational companion diagnostics. Instead of evaluating a single IHC assay, regulators may need to assess the complete system - from tissue preparation and staining to scanning, image management, algorithmic analysis, display, and the final biomarker report.Finally, we discuss what laboratories will need to implement these workflows, including validated scanning infrastructure, cybersecurity, tighter preanalytical process control, and training that helps pathologists interpret continuous computational measurements.Episode Highlights00:00 — Pathologists remain central to computational CDxWhy computational tools provide more precise measurements without replacing pathology expertise.01:09 — Why companion diagnostics are changingVisual IHC scoring helped launch precision oncology, but the model is approaching its limits.04:53 — The current companion diagnostic landscapeHow IHC, next-generation sequencing, liquid biopsy, and visual biomarker scoring are used today.07:01 — The mathematical burden placed on the human eyeWhy manually assessing tens of thousands of tumor cells requires pathologists to estimate rather than calculate.08:39 — The borderline patient dilemmaA digital tool can distinguish measurements such as 74% and 76%, while that difference is difficult to reproduce visually.09:27 — Why spatial context mattersComputational pathology can measure biomarker heterogeneity, clustering, and relationships between tumor and immune cells.12:17 — Where manual scoring reaches its limitsInterobserver variability, intraobserver variability, fatigue, staining interpretation, and heterogeneous tumors.18:29 — Moving from judgment calls to quantified measurementsWhy the next stage of precision oncology requires information beyond human visual perception.19:14 — Computer-assisted scoring versus computational CDxThe important distinction between helping a pathologist calculate an existing score and generating a new algorithmic measurement.23:22 — Computational pathology and decentralized workflowsHow digital images can support remote review, access to expertise, and second opinions.27:48 — Why therapies require higher-resolution biomarkersModern targeted treatments may respond to biological differences that categorical scoring can’t capture.32:09 — TROP2 in advanced non-small cell lung cancerThe episode’s example of a biomarker requiring computational measurement.33:26 — Understanding the normalized membrane ratioHow the algorithm measures membrane expression relative to total protein expression at the individual-cell level.35:46 — Working with regulators on a new diagnostic modelPurvi discusses global health authority engagement and the FDA Breakthrough Device Designation.38:33 — The computational CDx as a system of systemsWhy staining, scanning, image management, algorithms, displays, and reporting must be evaluated together.40:11 — Changes to validated workflow componentsHow using a different ...
    Show More Show Less
    1 hr and 7 mins
  • 245: Why Going Slow Is Killing Digital Pathology Adoption | Syed T. Hoda, M.D.

    245: Why Going Slow Is Killing Digital Pathology Adoption | Syed T. Hoda, M.D.
    Jul 29 2026
    Send us Fan MailIs your digital pathology rollout moving so slowly that it’s creating a fragmented workflow instead of transforming the department?In this episode of the Digital Pathology Podcast, I speak with Dr. Syed Hoda, Director of Digital Pathology at NYU, about why gradual implementation may no longer be the best approach to digital pathology adoption.Dr. Hoda explains how NYU used an intensive nine-month planning period to prepare for a department-wide transition. The process involved pathology, IT, project managers, vendors, hospital leadership, and approximately 40–50 people participating in regular planning calls.This wasn’t simply a scanner installation.The team mapped workflows, configured Epic Beaker, redesigned laboratory spaces, tested integrations, planned training, and addressed the practical concerns of nearly 100 pathologists.We also discuss why scanner specifications may matter less than integration, vendor support, training, and system performance. For Dr. Hoda, digital pathology had to work as smoothly as glass microscopy. Speed was non-negotiable.Change management played an equally important role. Through open discussions, town halls, and the ADKAR framework, the team addressed concerns ranging from ergonomics to the loss of collaborative microscope sessions.The result? Every pathologist adopted the digital workflow, no one left the department because of the transition, and approximately 60–65 pathologists now work remotely using equipment that matches their office setup.Finally, we examine the next step: artificial intelligence in pathology. Dr. Hoda explains why NYU focused on building a reliable digital foundation before introducing AI. He also raises important questions about validation, transparency, responsibility, regulatory clearance, and the need for greater pathologist involvement in AI development.Episode Highlights00:00 — Are we repeating the same mistakes with pathology AI?Dr. Hoda compares the current excitement around AI with the early promises made about digital pathology 15 years ago.01:04 — Meet Dr. Syed HodaHis clinical pathology background and path to becoming NYU’s Director of Digital Pathology.03:16 — Why going slowly can hold departments backHow partial adoption creates fragmented workflows, inconsistent training, and prolonged implementation.06:25 — Leadership support for rapid adoptionWhy institutional commitment, resources, and an ambitious timeline made the project possible.10:13 — Nine months of detailed planningWorkflow mapping, laboratory changes, system configuration, vendor selection, testing, and validation.11:48 — The role of professional project managementWhy pathologists shouldn’t be expected to coordinate every part of a complex digital transformation.14:29 — Why the scanner isn’t the most important decisionImage quality matters, but integration, service, training, and workflow fit may matter more.17:42 — People matter more than machinesHow vendor relationships and departmental engagement supported adoption.19:19 — Setting clear expectations across the departmentNYU communicated that every pathologist would move to digital sign-out within a defined period.20:49 — Change management is a structured processHow the ADKAR framework guided communication, education, adoption, and reinforcement.25:07 — Addressing practical and personal concernsFrom mouse ergonomics to preserving collaborative case review between pathologists.27:19 — Why NYU didn’t introduce AI firstDr. Hoda explains why pathologists needed to become comfortable with the digital platform before adding new AI tools.29:26 — Digital pathology and remote sign-outApproximately 60–65 pathologists now work remotely with equipment matching their office setup.30:28 — Why speed is non-negotiableEven a small delay or repeated pixelation can quickly undermine confidence in a digital workflow.33:25 — A cautious approach to pathology AIConcerns about premature adoption, self-validation, limited regulatory clearance, and lack of pathologist involvement.37:27 — Scientific validation, transparency, and responsibilityWhat happens when the AI result and the pathologist’s interpretation don’t agree?40:41 — Where AI could meaningfully augment pathologyQuantifying microenvironments, feature combinations, ratios, and findings that are difficult to assess visually.Resources MentionedADKAR change management frameworkDigital Pathology AssociationExecutive War CollegeFDA list of AI-powered medical devicesA radiology mock-trial paper examining responsibility when clinicians use AI: Examining perceptions of liability about AI in radiology (MedRxiv)Why AI cannot do good science without humans (Nature Editorial)A previous Digital Pathology Podcast discussion about AI-supported colorectal cancer feature analysis (How to use deep learning image analysis for colon cancer with Rish Pai)Listen to the full conversation for a practical look at digital pathology ...
    Show More Show Less
    45 mins
  • 244: Why AI Still Hasn't Revolutionized Drug Discovery (Yet) | Thibault Geoui, PhD

    244: Why AI Still Hasn't Revolutionized Drug Discovery (Yet) | Thibault Geoui, PhD
    Jul 22 2026
    Send us Fan MailIf AI is already being used across the drug development pipeline, why hasn’t its impact matched the investment?AI can help researchers review scientific literature, predict protein structures, prioritize molecules, assess toxicity, support clinical trials, and monitor adverse events. But access to better tools doesn’t automatically create better drugs.In this episode, I speak with Thibault Geoui, Science CDO and host of the Tech & Drugs Podcast, about where AI is making a practical difference in drug discovery and development—and where the results remain limited. We map AI across the full drug development funnel, from basic research and target identification to preclinical testing, clinical trials, regulatory documentation, commercialization, and pharmacovigilance.We also discuss why digital-native tech-bio companies may be better positioned to benefit from AI than traditional pharmaceutical organizations. The difference isn’t simply the model. It’s how data, people, laboratory experiments, and AI tools are connected inside the workflow.For digital pathology professionals, the conversation becomes especially relevant when we examine AI-powered biomarker development, the role of pathology in pharmaceutical research, and the Roche–PathAI case discussed in the episode.And, of course, we talk about the problem every AI user eventually faces: an answer can look polished, specific, and completely convincing—and still be wrong.Episode Highlights00:00 — When convincing AI output creates more work Why AI can accelerate information generation while increasing the time required for review and verification.02:15 — From structural biology to science and technology leadership Thibault shares his background in X-ray crystallography, structural biology, scientific data, and digital product development.15:36 — Understanding the drug discovery and development funnel How thousands of potential compounds are narrowed down through discovery, preclinical research, clinical trials, and approval.20:00 — AI for scientific literature review How alerts, filtering, summarization, and information extraction can help researchers manage a rapidly growing scientific literature base.22:32 — AlphaFold and protein structure prediction What faster access to predicted protein structures changes for researchers—and why structural prediction alone doesn’t solve drug discovery.24:13 — Searching an enormous chemical space How AI can help design and prioritize potential molecules for synthesis and experimental testing.25:50 — Predicting efficacy and toxicity Where AI supports preclinical research, why the models remain imperfect, and why experimental validation still matters.29:38 — Has AI changed drug development outcomes yet? A practical discussion about drug approval rates, AI investment, uneven returns, and the difference between deploying a tool and integrating it into a process.34:33 — Why traditional pharma struggles to scale AI Siloed data, legacy systems, organizational complexity, and the need to build reusable data workflows.37:57 — The “lab in the loop” model How tech-bio companies connect AI predictions with wet-lab experiments and feed the new data back into their models.44:37 — Can tech-bio companies shorten development timelines? How digital-native organizations are changing parts of the discovery and preclinical process.58:00 — AI, pharma, and digital pathology What the Roche–PathAI case discussed in the episode may indicate about the role of pathology data, biomarker discovery, and pharmaceutical workflows.01:06:17 — AI errors in regulated environments Why responsibility remains with the person or company submitting AI-assisted work, regardless of which tool produced it.01:17:37 — The growing cost of AI tools Subscriptions, token limits, model selection, AI orchestrators, and the need to use expensive tools more intentionally.01:27:50 — What successful AI adoption requires Starting with focused pilots, training scientists and technologists together, and treating implementation as organizational change.01:30:26 — The AI quirks that still frustrate users Hallucinated information, ignored writing instructions, stylistic habits, and poor awareness of time and context.The episode’s timestamped themes and examples are documented in the supplied summary. The broader discussion covers AI from literature mining and molecular design through clinical development and post-market monitoring. Resources Mentioned Thibault Geoui’s LinkedIn profile Tech & Drugs PodcastMIT NANDA study on generative AI implementation and return on investment Insilico Medicine as an example of a digital-native tech-bio company AI is already changing how scientific work gets done. The bigger question is whether organizations can redesign their workflows, train their teams, and maintain the human oversight needed to use it well.Listen to the full episode for a practical look at AI in...
    Show More Show Less
    1 hr and 37 mins
  • 243: How to Teach AI to Healthcare Professionals | Podcast with Candice Chu

    243: How to Teach AI to Healthcare Professionals | Podcast with Candice Chu
    Jul 2 2026
    Send us Fan MailWhat does AI literacy actually look like for pathologists, researchers, and future clinicians? And how do you teach it in a way that is practical, not abstract?In this episode, I talk with Candice Chu, DVM, PhD about something I think a lot of people in digital pathology and computational pathology are feeling right now: AI is moving fast, but education is still catching up.Candice is a clinical pathologist, veterinarian, and educator building AI-focused teaching and research at Texas A&M. We worked together before on digital pathology and image analysis projects, so this conversation felt especially grounded. We talk about her AI literacy curriculum framework for veterinary education, why she decided to build it, and what it takes to teach AI in a way that is useful, ethical, and realistic.This episode is about understanding what AI tools are good for, where they can waste your time, and why hands-on experience matters. Candice explains why she sees AI as a set of tools, not a belief system. Try them. Learn them. Keep what improves your workflow. Drop what does not.We also talk about the difference between putting educational content online and building formal institutional teaching. That matters because social media can move quickly, but curriculum changes, research, and professional organizations shape longer-term adoption. Candice shares how her course started as a low-stakes elective, then grew into a more structured framework that combines education with publishable research.A big part of this conversation is the curriculum itself. We go through what students actually learn: AI fundamentals without heavy math, machine learning and image analysis, large language models, prompt engineering, chatbot building, ethics, literature research, and final projects where students evaluate real tools and workflows. I liked that the course does not stop at theory. It asks students to use tools, question them, and explain where they help and where they do not.We also get into something that matters far beyond veterinary medicine: professional responsibility. If AI is involved in a workflow, the clinician is still responsible. That includes fabricated citations, bad outputs, weak prompts, and the temptation to trust tools too quickly. Candice makes a strong case that AI education needs ethics, legal context, and interdisciplinary teaching built in from the start.If you are trying to think more clearly about AI in pathology, education, workflow design, or professional training, this episode gives you a concrete example of what responsible AI literacy can look like.Episode Highlights00:00 – Why AI tools are just tools, and why trying them matters even if you later decide not to keep using them00:33 – Who Candice Chu is and why her work on AI literacy in veterinary medicine is worth paying attention to02:33 – Why going back to Texas A&M changed the scale of Candice’s AI research and teaching07:53 – How the AI course was designed as a low-stakes elective first, and why that helped student engagement11:16 – Where veterinary AI education stands now, and what professional organizations like ACVP are doing13:08 – Why AI adoption in veterinary medicine is still slow, and what skepticism usually sounds like in practice15:19 – Real examples of how Candice uses LLMs and computer vision in pathology, medical records, and research19:58 – What is actually inside the 15-week AI literacy curriculum, from fundamentals to final projects24:16 – Why ethics and legal responsibility are not optional in AI education31:35 – Why no-code tools and vibe coding are entering the curriculum already38:50 – The AI tools Candice is testing in her own workflow, including Claude, Codex, and PerplexityResources mentionedCandice Chu’s AI literacy curriculum framework paper in Frontiers in Veterinary ScienceCandice’s earlier work on ChatGPT in veterinary medicineTexas A&M and the institutional setting where Candice is building AI research and teachingMr. Don Riddick and the AVMA AI working group, mentioned in the ethics and legal contextClaude, Codex, and Perplexity as AI tools Candice is actively testingDigital Pathology 101, mentioned in the conversation as a teaching resourceCandice’s online educational work on Instagram. Support the showGet the "Digital Pathology 101" FREE E-book and join us!
    Show More Show Less
    44 mins
  • 242: Foundation Models in Pathology: Strong on Paper, Ready for Labs?

    242: Foundation Models in Pathology: Strong on Paper, Ready for Labs?
    Jun 24 2026
    Send us Fan MailAre pathology foundation models actually ready for labs, or are they still stronger on paper than in practice?In this episode of DigiPath Digest #49, I unpack a timely review on pathology foundation models and ask the question that matters most to me: not just what these models can do, but what has to be true before they are genuinely useful in real pathology workflows.I walk through how pathology AI moved from narrow, task-specific models into the era of transformer-based foundation models. That shift matters because pathology is no longer only about looking at H&E in isolation. Today, pathologists are expected to integrate morphology, immunohistochemistry, molecular assays, genomics, and clinical context. That growing complexity is one reason foundation models are getting so much attention.In this discussion, I explain how transformers entered pathology, why image patches are treated like tokens, and how shared embeddings can support classification, regression, segmentation, and multimodal retrieval. I also go through the major pathology foundation models mentioned in the paper, including Virchow/Virchow2, Mayo Clinic Atlas, UNI, CONCH, H-Optimus, GigaPath, and TITAN, and why scale alone is not the full story.A big part of this episode is about the gap between benchmark performance and clinical readiness. I talk about the persistent limitations in training data diversity, the overuse of TCGA, and why public benchmarks can still miss what real pathology practice looks like. I also cover where foundation models still struggle, especially in cytopathology, hematopathology, and underrepresented disease areas, along with the real-world problems of artifacts, domain shift, concept drift, infrastructure burden, regulatory complexity, and workflow disruption.For me, one of the most important themes is this: AI in pathology should augment, not replace, pathologists. The future is not about handing diagnosis to a model. It is about building tools that support pathologists better, fit real workflows, and can be validated in ways that deserve trust.I also spend time on what comes next: explainable AI, counterfactual explanations, conversational interfaces, retrieval-augmented systems, multimodal fusion, and the need for deployment-centric validation rather than paper-only excitement.If you are trying to understand where pathology foundation models really stand today, this episode will help you separate the promise from the practical barriers.Episode Highlights00:01 – Why I chose this paper, what is changing at Digital Pathology Place, and why foundation models are worth paying attention to now.02:15 – The core questions: what pathology foundation models are, where they are, and how difficult they are to apply in pathology.04:50 – Why pathology is becoming more cognitively demanding, and how multimodal complexity is driving interest in scalable AI.07:02 – From narrow AI to transformers: how pathology moved beyond single-task CNN models.10:16 – How transformers work in pathology: image patches as tokens, self-attention, embeddings, and downstream tasks.14:16 – Why multimodality matters, and what kinds of data foundation models may eventually integrate.15:27 – Timeline of key model developments, from “Attention Is All You Need” to gigapixel-scale pathology foundation models.17:13 – The leading models and what scale really looks like: Virchow, Mayo Clinic Atlas, UNI, CONCH, H-Optimus, and GigaPath.19:51 – Why dataset diversity matters more than sheer volume, and why TCGA is not enough.23:17 – Where foundation models still struggle: cytopathology, hematopathology, rare disease, artifacts, scanner shifts, and pen marks.28:06 – Explainability, counterfactual explanations, and why trust in pathology AI needs more than attention maps.30:17 – The real deployment hurdles: regulation, infrastructure, workflow fit, and economics.36:32 – Why AI should augment pathologists, not replace them, and which tedious tasks pathologists would gladly hand over.38:36 – Retrieval-augmented and conversational AI in pathology: where interactive systems may actually help.40:51 – Vision-language models and multimodal fusion with histology, radiology, genomics, and clinical notes.42:16 – The path forward: deployment-centric design, prospective multi-site validation, and human-AI collaboration.44:08 – Closing thoughts on AI literacy, community learning, and what needs to happen next.Resources MentionedMain paper discussed:Pathology Foundation Models: Evolution, Current Landscape, Challenges and Opportunities from a Technical and Clinical Perspectivehttps://doi.org/10.3390/bioengineering13050577Review article / journal landing page:https://doi.org/10.3390/bioengineering13050577Benchmarks mentioned:PathoBench — discussed in the review paper; use the review link here for context until you want to swap in a canonical project page:https://doi.org/10.3390/bioengineering13050577PathBench — ...
    Show More Show Less
    44 mins