Thank you for this amazingly well thought piece! Having worked in a big pharma in a pathology group (as a scientist), I understand the caveats of switching to digital pathology entirely. The workflow is often distributed between pathologists, scientists, computational scientists and AI/ path softwares with all of them bringing their own specialization and contributions to the process. AI path softwares definitely contribute a lot in streamlining the process and making the lives of scientists/ pathologists easier. We are far from automating the entire workflow (after the staining and imaging process) but I would still go for the product based approach that you mentioned above. Products from Indica labs (like HALO AI) or even imaging solutions from Akoya help in reducing the time to process thousands of slides in a day. Even introducing small tools/ features as simple as image markups/ AI based image classifiers, reduces the TAT that can contribute a lot in a fast paced environment. One approach would be understanding the workflow in detail and brainstorming what small (but significant) features can be introduced.
One key benefit for product based approach is the market increases exponentially. Products can be marketed even in other countries where the market is still not saturated by labcorp/ Quest or other companies mentioned above. Service based approach remains limited to logistics and procurement if the slides need to be physically shipped to the location. I am not a product expert but a well wisher who thinks that these advanced tools will benefit patients from other parts of the world where they are currently not available. How much investment the institutions/ pharma companies/ hospitals are willing to make outside of US is another question but I am positive that there will be investment incentives in a wider market nonetheless.
Just a note of thanks for a well-written, and well-thought-out piece. Plenty of failures and plenty of notable new entrants to me signals a healthy future, if one or more of these can crack the code and get lucky (i.e. the right technology with the right niche at the right time = luck).
In the late 2010's I worked for a startup whose founder also had an ML/Big Data tool to dig through California's state healthcare system (Kaiser Permanente's) EHR to identify anomalies. It found quantifiable savings - yet strangely enough not enough 'economic value' to see that business expand the way I thought it would (from a bystander's point of view).
Lastly, at last year's AACR I learned in an AI plenary that Europe has already approved an AI system that can determine MSI status from H&E stains, which I thought was remarkable, to some 60-70% accuracy. (Sorry no more details on that, I'm just an observer...) Keep up the nice work!
Pathologist here (who just sat through a PathAI pitch). Agree with all of this but some US institutions have made it work and have gone fully digital (e.g Mayo Rochester). People at those places don't want to practice elsewhere. Cases aren't lost, you can sign out anywhere, old cases are readily available for comparison, distributed institutions are supported. The expense of maintaining a glass filing system is not trivial either. Slides are not thrown out after review (that was the one odd comment) but filed. AI only companion diagnostics are coming which makes the case stronger. When cases can be diagnosed and reports generated automatically, it does save overall time. It is true that the system has to be part of a workflow and not a stand alone product.
This analysis accurately reflects what many of us in the field have long suspected: the promise of AI in pathology hasn’t failed—but its maturation is unfolding far more slowly than we anticipated. Not due to a lack of technology, but because of a combination of structural, economic, and cultural factors deeply rooted in clinical practice.
It’s true that digitization transformed radiology decades ago, but in pathology, workflows remain predominantly manual, fast, and built around the expertise of the pathologist. For an AI tool to be adopted, it’s not enough to be “accurate”—it must integrate seamlessly, offer tangible clinical value, and, most importantly, not slow things down.
The dilemma between service-based and product-based business models is real. I’ve seen brilliant solutions remain stuck in perpetual pilot phases because they didn’t fit within institutional structures, and others get dismissed because the value added was not enough to justify the effort of adoption.
What I found most insightful in this essay is the conclusion: the future isn’t about replicating the pathologist, but expanding what the pathologist can do. Tools like Artera’s (providing prognostic and predictive insights directly from biopsies) or Pictor Labs’ virtual staining technology point to novel clinical applications that wouldn’t be possible without AI.
As I see it, we’re moving beyond the “AI for hype” phase and entering a new era of specialization, where success will come from solving specific, meaningful problems—not from promising universal solutions. And while that may be less flashy, it’s a sign of maturity that gives me hope.
Thank you for this amazingly well thought piece! Having worked in a big pharma in a pathology group (as a scientist), I understand the caveats of switching to digital pathology entirely. The workflow is often distributed between pathologists, scientists, computational scientists and AI/ path softwares with all of them bringing their own specialization and contributions to the process. AI path softwares definitely contribute a lot in streamlining the process and making the lives of scientists/ pathologists easier. We are far from automating the entire workflow (after the staining and imaging process) but I would still go for the product based approach that you mentioned above. Products from Indica labs (like HALO AI) or even imaging solutions from Akoya help in reducing the time to process thousands of slides in a day. Even introducing small tools/ features as simple as image markups/ AI based image classifiers, reduces the TAT that can contribute a lot in a fast paced environment. One approach would be understanding the workflow in detail and brainstorming what small (but significant) features can be introduced.
One key benefit for product based approach is the market increases exponentially. Products can be marketed even in other countries where the market is still not saturated by labcorp/ Quest or other companies mentioned above. Service based approach remains limited to logistics and procurement if the slides need to be physically shipped to the location. I am not a product expert but a well wisher who thinks that these advanced tools will benefit patients from other parts of the world where they are currently not available. How much investment the institutions/ pharma companies/ hospitals are willing to make outside of US is another question but I am positive that there will be investment incentives in a wider market nonetheless.
Just a note of thanks for a well-written, and well-thought-out piece. Plenty of failures and plenty of notable new entrants to me signals a healthy future, if one or more of these can crack the code and get lucky (i.e. the right technology with the right niche at the right time = luck).
In the late 2010's I worked for a startup whose founder also had an ML/Big Data tool to dig through California's state healthcare system (Kaiser Permanente's) EHR to identify anomalies. It found quantifiable savings - yet strangely enough not enough 'economic value' to see that business expand the way I thought it would (from a bystander's point of view).
Lastly, at last year's AACR I learned in an AI plenary that Europe has already approved an AI system that can determine MSI status from H&E stains, which I thought was remarkable, to some 60-70% accuracy. (Sorry no more details on that, I'm just an observer...) Keep up the nice work!
Pathologist here (who just sat through a PathAI pitch). Agree with all of this but some US institutions have made it work and have gone fully digital (e.g Mayo Rochester). People at those places don't want to practice elsewhere. Cases aren't lost, you can sign out anywhere, old cases are readily available for comparison, distributed institutions are supported. The expense of maintaining a glass filing system is not trivial either. Slides are not thrown out after review (that was the one odd comment) but filed. AI only companion diagnostics are coming which makes the case stronger. When cases can be diagnosed and reports generated automatically, it does save overall time. It is true that the system has to be part of a workflow and not a stand alone product.
Thank you for reading and for the information! Removed the 'thrown out' bit :)
I know a very well-positioned founder working on launching something quite exciting in this space, if anyone reading this is curious for an intro
Sure, what’s that?
If you DM me your email, I can connect you
This analysis accurately reflects what many of us in the field have long suspected: the promise of AI in pathology hasn’t failed—but its maturation is unfolding far more slowly than we anticipated. Not due to a lack of technology, but because of a combination of structural, economic, and cultural factors deeply rooted in clinical practice.
It’s true that digitization transformed radiology decades ago, but in pathology, workflows remain predominantly manual, fast, and built around the expertise of the pathologist. For an AI tool to be adopted, it’s not enough to be “accurate”—it must integrate seamlessly, offer tangible clinical value, and, most importantly, not slow things down.
The dilemma between service-based and product-based business models is real. I’ve seen brilliant solutions remain stuck in perpetual pilot phases because they didn’t fit within institutional structures, and others get dismissed because the value added was not enough to justify the effort of adoption.
What I found most insightful in this essay is the conclusion: the future isn’t about replicating the pathologist, but expanding what the pathologist can do. Tools like Artera’s (providing prognostic and predictive insights directly from biopsies) or Pictor Labs’ virtual staining technology point to novel clinical applications that wouldn’t be possible without AI.
As I see it, we’re moving beyond the “AI for hype” phase and entering a new era of specialization, where success will come from solving specific, meaningful problems—not from promising universal solutions. And while that may be less flashy, it’s a sign of maturity that gives me hope.
And what about China? It's the 800lb gorilla of ML.
The Paige.AI story is also interesting due to some high profile controversies with their relationship to MSKCC and some recent cofounder exits.