Digital Pathology Podcast

Send us Fan Mail
Are 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 Highlights
00:01 – 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 Mentioned
Main paper discussed:
Pathology Foundation Models: Evolution, Current Landscape, Challenges and Opportunities from a Technical and Clinical Perspective
https://doi.org/10.3390/bioengineering13050577
Review article / journal landing page:
https://doi.org/10.3390/bioengineering13050577
Benchmarks 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/bioengineering13050577
PathBench — public benchmark paper:
https://arxiv.org/abs/2505.20202
MEDFAIR — benchmark paper:
https://arxiv.org/abs/2210.01725
MEDFAIR code repository:
https://github.com/ys-zong/MEDFAIR

Models mentioned:Model overview in the review (Virchow/Virchow2, UNI, CONCH, H-Optimus, GigaPath, TITAN, Mayo Clinic Atlas):
https://doi.org/10.3390/bioengineering13050577
Virchow:
https://arxiv.org/abs/2309.07778
UNI:
https://arxiv.org/abs/2308.15474
CONCH:
https://arxiv.org/abs/2307.12914
Mayo Clinic Atlas:
https://arxiv.org/abs/2501.05409
TITAN:
https://arxiv.org/abs/2411.19666

Dataset mentioned:
The Cancer Genome Atlas (TCGA)
https://portal.gdc.cancer.gov/
Book mentioned:
Digital Pathology 101: All You Need to Know to Start and Continue Your Digital Pathology Journey
https://digitalpathologyplace.com/
Platform:
Digital Pathology Place
https://digitalpathologyplace.com/
Support the show
Get the "Digital Pathology 101" FREE E-book and join us!

Send us Fan Mail
How far can pathologists take visual biomarker scoring before human vision becomes the bottleneck?
In this episode, I talk with Doug Bowman. PhD, VP Precision Medicine at Indica Labs, about what happens when companion diagnostics move from traditional visual scoring into the era of AI-powered image analysis. Doug comes from a biomedical and electrical engineering background, with experience in microscopy, digital image analysis, pharma workflows, and now precision medicine at Indica Labs. That combination makes him a great person to talk to about how image analysis actually fits into real companion diagnostic development.
We start with a very practical question: what is a companion diagnostic, and why is it becoming so important in precision medicine? Doug explains that companion diagnostics are developed alongside therapeutics to help identify which patients are most likely to benefit from a specific treatment, especially in more complex therapies like antibody-drug conjugates (ADCs). We use HER2 as an example, and from there we get into the real challenge: once a biomarker cutoff matters clinically, visual estimation around that cutoff becomes much harder than many people want to admit.
That is where this conversation gets especially useful for pathologists and digital pathology trailblazers. We talk about the limits of human vision, why low or ultra-low biomarker expression is difficult to score consistently, and how AI helps at multiple levels of the workflow: slide QC, tissue classification, cell segmentation, membrane and cytoplasmic measurement, and spatial analysis. Doug makes the case that AI is not only a convenience here. In some cases, it is the only realistic way to capture the kind of quantitative information modern therapies need.
We also get into one of the more interesting examples from the episode: the Trop2 story, where a ratio of cytoplasmic to membrane expression appears to predict therapeutic efficacy better than looking at one compartment alone. That kind of compartment-level quantitation is exactly where computational pathology becomes more than a digital version of what the eye already does. It starts uncovering measurements and signatures the eye cannot reliably extract on its own.
Another important part of the discussion is workflow and regulation. Doug walks through how AI-powered companion diagnostics are developed from preclinical work, to human feasibility studies, to RUO or clinical trial assays, and eventually toward analytical and clinical validation with regulatory engagement happening early. We also talk about the Indica Labs and Leica Biosystems partnership, and why end-to-end capability matters when you are trying to build something clinically deployable rather than just analytically interesting.
What I liked about this conversation is that it stayed grounded. We did not talk about AI as magic. We talked about image analysis as a method, companion diagnostics as a workflow, and precision medicine as something that only works when the measurement is good enough to support real decisions.

Episode Highlights
00:00 – Why AI matters in slide QC, tissue classification, and cell-level analysis before you even get to the biomarker score.
00:54 – Doug Bowman’s background in biomedical engineering, microscopy, and digital image analysis.
05:16 – What a companion diagnostic actually is, and why it is critical for targeted therapies and ADCs.
07:34 – Why visual biomarker scoring becomes unreliable around critical cutoffs, especially in low-expression cases.
10:09 – How AI expands the workflow: slide QC, tissue classification, and precise cell segmentation.
13:07 – Why pathologists remain central in AI workflows through validation, markup review, and model refinement.
16:31 – The Trop2 example: when cytoplasmic-to-membrane ratio tells you more than one compartment alone.
20:23 – The Indica Labs + Leica Biosystems partnership and why end-to-end workflow matters in companion diagnostics.
22:53 – What the development journey looks like from early algorithm work to RUO, validation, and regulatory interaction.
26:51 – Multiplexing, spatial analysis, and why more clinical value often comes with more deployment complexity.
33:29 – Why image analysis literacy matters, and how shared language between pathologists and scientists becomes essential.
40:13 – Where to learn more about Indica Labs and who to contact for collaboration.

Resources mentioned
Indica Labs 
Indica Labs contact – [email protected]
HALO software / HALO AI diagnostic image analysis – discussed in the context of companion diagnostic deployment and pharma services.
Leica Biosystems GT450DX – referenced as an FDA-cleared slide scanner in the Indica-Leica partnership.
Digital Pathology Association – mentioned as part of the broader educational ecosystem for digital pathology and image analysis.
Digital Pathology Place / Digital Pathology Podcast – the platform hosting this conversation and related education around digital pathology and AI.
Support the show
Get the "Digital Pathology 101" FREE E-book and join us!

Send us Fan Mail
Can AI copilots really keep up with pathologists when the cases are new, the workflow is messy, and the benchmark is actually protected from leakage?
In this episode of DigiPath Digest #48, I focus on one paper: DALPHIN: Benchmarking Digital Pathology AI Copilots Against Pathologists on an Open Multicentric Dataset. I chose this paper because I think the field needs more of this kind of work. Less hype. More evaluation. Less “look what AI can do.” More “how do we test it in a way that actually means something?” 
In this session, I look at what makes DALPHIN important for pathologists, lab leaders, and digital pathology trailblazers trying to make sense of pathology AI right now. The paper benchmarks three models against human pathologists: two general-purpose models, Gemini 2.5 Pro and GPT-5, and one pathology-specific model, PathChat+. The dataset includes 1,236 images from 300 cases, covering 130 diagnoses, 14 pathology subspecialties, and cases from six countries. Human performance is benchmarked with 31 pathologists from 10 countries. 
What I like about this paper is that it does not stop at top-line performance. It deals with the benchmarking problem itself. The authors built a sequestered, indirectly accessible ground truth so the evaluation data could not simply be scraped into model training. That matters because without that protection, benchmarking can become an illusion of genius rather than a real test of generalization. 
The results are interesting and more nuanced than a simple win-or-lose story. PathChat+ reached expert-level performance in four of six tasks, Gemini in two of six, and GPT in one of six. That tells us something important already: pathology-specific training matters. But it also does not mean pathology is solved. In organ recognition, expert pathologists still outperformed all the models. In rare cancers, none of the models reached expert-level performance. And in ambiguous cases, the models still struggled with something human pathologists do all the time: expressing uncertainty. 
I also spend time on one of the most practical parts of the paper: model behavior. Gemini tended to overcall. GPT tended to undercall. PathChat was more balanced. That matters in practice. A pathologist using a copilot needs to know the tool’s calibration bias before they can safely interpret what it is telling them. I also talk about anchoring bias in conversational interfaces, where early hallucinations can propagate through later answers if memory is not reset between questions. That is not just a technical curiosity. That is a workflow and safety issue. 
Why should you listen? Because this episode is really about a bigger question: What kind of evidence should pathologists demand before AI copilots enter real workflows? If you want to understand validation, data leakage, rare-case performance, uncertainty, and why these tools should still be treated as co-pilots rather than autopilots, this is a useful paper to know. 
Episode Highlights
01:20 – Why I chose the DALPHIN preprint and why benchmarking matters right now. 
05:38 – What is in the DALPHIN dataset: 300 cases, 130 diagnoses, 14 subspecialties, 6 countries. 
07:57 – Top-line performance: PathChat+ reaches expert-level performance in 4 of 6 tasks. 
09:41 – The benchmarking trap of data leakage and why DALPHIN’s sequestered ground truth matters. 
12:19 – Why real pathology diagnosis is not text-only and why macro + micro context matters. 
15:26 – Tissue recognition, neoplasm detection, ambiguity, and conversational memory: how the testing was structured. 
21:29 – The diagnostic personalities of the models: overcalling, undercalling, and balanced behavior. 
24:36 – Rare cancers: where AI copilots still fall short of expert human performance. 
28:00 – Why binary outputs are not enough when pathology often lives in uncertainty. 
31:37 – Anchoring bias and conversational memory: how early hallucinations can keep propagating. 
37:11 – Why these tools should be treated as co-pilots, not autopilots. 
40:29 – Resources for beginners: Digital Pathology 101 and continued AI literacy. 
Resources mentioned
DALPHIN preprint: arXiv:2605.03544v1 
DALPHIN evaluation platform: dalphin.grand-challenge.org 
PathChat+ pathology-specific AI model discussed in the benchmark. 
Digital Pathology 101 free eBook by Dr. Aleksandra Zuraw. 
Educational streams on tissue recognition and computer vision literacy mentioned in the session.
Support the show
Get the "Digital Pathology 101" FREE E-book and join us!

Send us Fan Mail
Do you really need a scanner, whole slide images, and AI infrastructure before you can start in digital pathology?
In this episode, I argue that you do not.
I’m Dr. Aleksandra Zuraw, veterinary pathologist and digital pathology educator, and this talk is about a belief I hear all the time: I don’t have the tools yet, so there is no point learning digital pathology. I used to think that too. When I was training in Berlin, there was one Leica 6-slide scanner, and it felt like digital pathology was only for a small group of chosen people. That experience made the field feel distant, exclusive, and not really available to beginners. 
What changed for me was not a new scanner. It was a small project.
I needed a more consistent way to quantify a senescence marker in archived skin samples, so I used a microscope camera, captured images, opened them in Microsoft Paint, and manually marked cells with colored dots. It was scrappy. Very low tech. But it was also digital, consistent, and verifiable. That project became my first real step into digital pathology and helped me get my first job in the field, where I worked between pathologists and image analysis scientists on biomarker quantification and patient stratification problems. 
That is the core point of this episode: knowledge unlocks technology.
Scanners matter. AI tools matter. But the deeper bottleneck is whether enough people understand how to use these tools, ask good questions, and connect pathology expertise with digital workflows. That is why this episode is really about readiness. Not readiness of the hardware. Readiness of the people.
I also talk about Dr. Taladzer from Pakistan, whose story makes this point even more clearly. At the time, Pakistan had around 220 million people, about 500 pathologists, and zero scanners. She still started learning digital pathology during COVID using a microscope and camera, joined the Digital Pathology Association, taught herself from papers and online resources, and kept going even after multiple AI vendors rejected her because she did not have whole slide images. Eventually, she found a DIY image analysis platform, learned to annotate and train models on static images, completed projects quickly, and went on to publish more than 10 digital pathology papers without ever using WSI.
Why should you listen?
Because this episode is for pathologists and lab leaders who are interested in digital pathology but still feel stuck at the beginning. It is for people waiting for permission, perfect infrastructure, or a formal roadmap. And it is for trailblazers who came back from a meeting or conference energized, but need a practical way to turn that energy into action before it fades.
I also address an important AI question near the end: How do we know an AI model is good enough for pathology? I talk about why models are only as good as the pathologist annotations used to train them, why concordance between pathologists matters, how orthogonal labels like IHC can improve model quality, and why pathologists still need to stay in the loop as these systems develop and get deployed.
If you are trying to figure out where to start, this episode gives you a practical answer: start where you are. Start with what you have. Start learning now.
Episode Highlights
00:00 – Why the real barrier to digital pathology is usually not the hardware
00:33 – What it feels like to be at the beginning of the digital pathology journey
02:50 – My first practical digital pathology project using a microscope camera and Microsoft Paint
05:37 – How that low-tech project led to my first digital pathology job
08:52 – Why knowledge, not infrastructure, is the real unlock
09:57 – Dr. Taladzer’s story: starting digital pathology in Pakistan with zero scanners
12:03 – What happened after repeated vendor rejection and why persistence mattered
14:39 – The “forgetting loop” vs the “commitment loop” after conferences
16:48 – Practical next steps: book, PubMed alerts, journal clubs, webinars, vendor resources
18:52 – Why I believe digital pathology is the gateway to faster diagnosis
20:00 – How to think about whether an AI model is really ready for pathology
Resources Mentioned
Digital Pathology 101 – free book recommended as a starting point for learning digital pathology. 
Digital Pathology Association – mentioned as a learning resource and professional community. 
PubMed alerts for AI and digital pathology. 
Journal clubs – mentioned as one way to keep learning consistently. 
Webinars and vendor resources – suggested as practical ways to keep building knowledge. 
A4A – the DIY image analysis platform that supported Dr. Taladzer’s early work with static image annotation and model training. 
Support the show
Get the "Digital Pathology 101" FREE E-book and join us!

Send us Fan Mail
Why are pathology vendors still speaking different image languages when radiology solved that problem decades ago?
In this episode of DigiPath Digest #46, I talk through four papers that all point to a bigger issue in digital pathology: we are not only dealing with better algorithms. We are dealing with interoperability, workflow design, explainability, and whether the field is actually ready to use these tools well.
I start with DICOM in digital pathology, because I think this is still one of the most important infrastructure questions in the field. Digital pathology has clear value for consultation, image analysis, archival, and workflow, but vendor-specific whole slide image formats still create silos. In the episode, I explain why DICOM matters, why adoption is still low, how the multi-resolution pyramid works, and why this is really about enterprise imaging and future-proofing, not just file conversion. 
Then I move into kidney transplant rejection, where the paper makes a strong case for multimodal precision diagnostics. Creatinine is late. Antibody testing can miss important biology. Biopsies can miss the area that matters. So the opportunity is not to replace pathology, but to combine biomarkers, biopsy, and machine learning in a way that is more useful than any one signal alone. I also talk about explainability here, because if a model gives a risk score, we need to know what contributed to it. 
The third paper focuses on perineural invasion in solid tumors, and I liked this one a lot because it shows how AI can help standardize something that is clinically important but still inconsistently detected and reported. Perineural invasion is not just a passive pathway of spread. The biology is more active than that, and the quantification can go far beyond a simple yes-or-no answer. This is a good example of where digital pathology can do something humans cannot realistically do by eye at scale. 
The last paper is on gastric cancer immunohistochemistry biomarkers and advanced quantification, including HER2, PD-L1, mismatch repair, and CLDN18.2. This section is really about complexity. We are now asking pathologists to visually score biology that is getting harder and harder to summarize consistently, especially when markers, spatial context, and multiplexing all start to matter at once. I make the case that computational pathology is becoming necessary here, not because pathologists are failing, but because the biology is outgrowing purely visual workflows. 
What ties these four papers together is simple: digital pathology is not only about remote reading anymore. It is about interoperability, quantification, explainable AI, and making pathology more precise in places where the old workflow is reaching its limit. If you are a pathologist, lab leader, or digital pathology trailblazer trying to figure out what actually matters right now, this episode will help you connect the dots.
Episode Highlights
 07:41 – Why DICOM still matters if we want digital pathology systems to work together.
14:39 – Current adoption of SVS, MRXS, and DICOM, and why DICOM is still lagging.
16:44 – How the DICOM whole slide image pyramid works and why it matters for workflow.
24:29 – Why kidney transplant rejection is still difficult to diagnose with any single marker.
29:18 – Why perineural invasion is clinically important and still inconsistently reported.
34:44 – How AI can quantify tumor-nerve relationships more consistently than visual review alone.
46:39 – Why gastric cancer biomarker scoring is getting too complex for purely visual workflows.
54:55 – Multiplexing, spatial biology, and why explainable AI matters in biomarker interpretation.
01:04:01 – What is really blocking digital pathology adoption: cost, workflow, regulation, or mindset? 
Resources mentioned
DICOM / digital pathology interoperability paper
https://pubmed.ncbi.nlm.nih.gov/42093730/
Kidney transplant rejection, biomarkers, and artificial intelligence
https://pubmed.ncbi.nlm.nih.gov/42073482/
Perineural invasion in solid tumors with AI and machine learning applications
https://pubmed.ncbi.nlm.nih.gov/42100436/
Gastric cancer IHC biomarkers, advanced detection methods, and perspectives
https://pubmed.ncbi.nlm.nih.gov/42075555/
Digital Pathology Place
https://digitalpathologyplace.com
Digital Pathology 101
Free PDF book mentioned at the end of the episode through Digital Pathology Place.
Support the show
Get the "Digital Pathology 101" FREE E-book and join us!