206: AI Applications in Oral and Maxillofacial Pathology

Paper Discussed in this Episode:

Artificial Intelligence and Its Applications in Oral and Maxillofacial Pathology. Veremis B. Dent Clin North Am. 2026 Apr;70(2):403-416.

Episode Summary: In this Journal Club edition of the Digital Pathology Podcast, we explore a wild paradox at the bleeding edge of diagnostic medicine. We examine a 2026 paper on artificial intelligence in oral and maxillofacial pathology that reveals a fascinating reality: while highly advanced AI models can match human experts in detecting diseases, their clinical rollout is completely blocked by a surprisingly analog problem. We unpack why a 15-second difference in a laboratory dye bath might thwart billion-dollar neural networks and what this means for the future of the pathology lab.

In This Episode, We Cover:

• The Baseline - Matching Human Experts: How AI currently performs at human-expert levels for straightforward diagnostic tasks, such as detecting squamous cell carcinoma.

• The Predictive Frontier (Prognostication): How AI goes beyond binary diagnosis to evaluate complex spatial relationships—like calculating the precise micrometer distance between every single tumor-infiltrating lymphocyte and the invading edge of a carcinoma. We discuss the holy grail of predicting malignant transformation in oral premalignant disorders.

• The Analog Roadblock - Pre-analytical Variance: Why the physical, multi-step process of turning a tissue biopsy into a glass slide using H&E (hematoxylin and eosin) staining introduces massive data variability that severely confuses AI models.

• The "Mojave Desert" AI Trap: How human brains abstractly interpret a dark pink cell, while an AI algorithm sees a fundamentally different mathematical environment of numerical RGB pixel values. We discuss why an algorithm trained perfectly on one lab's specific slides will completely fail when fed slides from a different lab with slight chemical variations, much like a self-driving car trained in the desert crashing in a blizzard.

• The Data Drought: Why we desperately need millions of whole slide images from thousands of different laboratories to train robust, open-source AI models, and why these multi-institutional, standardized public datasets simply don't exist yet.

• The Ultimate Dilemma for Local Labs: Will the inevitable adoption of AI diagnostic tools force independent pathology labs to abandon their unique, decades-old tissue preparation methods in favor of a single, universally mandated global standard for tissue fixation and staining?.

Key Takeaway: The true bottleneck for AI in oral pathology isn't a lack of computational horsepower; it is analog inconsistency. Until the pathology field can standardize pre-analytical tissue preparation and build massive, publicly available datasets, highly sophisticated AI algorithms will remain isolated in the research lab instead of fulfilling their massive potential in everyday clinical diagnostics

Get the "Digital Pathology 101" FREE E-book and join us!

Transcript

Imagine building this like multibillion dollar supercomputing infrastructure,

right? Just massive scale.

Exactly. You've got the most advanced neural networks ever coded and they're designed to predict the exact future of a cancer cell with, you know, nearperfect accuracy.

A total modern medical marvel,

right? It's ready to deploy only to have the entire thing completely thwarted because a technician in a basement laboratory somewhere left a piece of glass in a bath of pink dye for like 15 seconds too long.

Yeah, it's kind of absurd when you put it like that, but that is exactly where we are.

It really is. Welcome back, Trailblazers, to another Journal Club edition of the Digital Pathology Podcast. Today, we're exploring this wild paradox at the bleeding edge of diagnostic medicine.

It's a great topic.

So, our mission today is a deep dive into a brand new 2026 paper. It's by Dr. Brandon Vermis, published in the dental clinics of North America, and it's titled artificial intelligence and its applications in oral and maxilloacial ology.

Yeah. And the central tension of this research is just fascinating because you know when we hear about AI and healthcare, the conversation is almost entirely dominated by clinical or radiologic applications,

right? Like an algorithm scanning an MRI for a brain tumor.

Exactly. Or uh flagging a cavity on a dental X-ray. Those modalities are already digitized natively. The machine understands them

because it's a digital picture to begin with.

Yes. But microscopic tissue examination taking a physical literal piece of a patient's mouth and turning it into a diagnosis. That is an entirely different beast.

Oh, completely.

We're attempting to force this highly analog, very physical tradition into a rigid digital framework.

And it pushes back.

Okay, let's unpack this because before we get into the physical roadblocks thwarting those, you know, billion-dollar algorithms, we have to establish the baseline. Like what is this technology actually doing right now in the lab,

right? What's the current state-of-the-art

exactly? And Vermis makes it really clear that we aren't talking about science fiction here. In the realm of oral and mexylacial pathology when it comes to detecting relatively straightforward diseases,

these AI models are already matching human baselines,

which is frankly a staggering computational achievement. Matching a board-certified human pathologist on foundational diagnostic tasks is huge.

Yeah. But that's just the baseline, right?

Right. Because the threshold for real utility isn't just basic dis disease recognition the real diagnostic power emerges when we look at how the AI handles uh architectural complexity

like what give me an example

okay take squirm cell carcinoma

okay

the algorithm isn't just making a binary you know cancer or no cancer decision it is actively evaluating the highly intricate spatial relationships between cells

oh wow so it's looking at the actual geometry

yes is mapping out the nuclear to cytoplasmic ratios across tens of thousands of cells simultaneously

that's in

and quantif ing hyperchromasia which is basically measuring exactly how dark the abnormal nuclei are staining compared to healthy tissue.

So it's way beyond simple pattern matching at this point. It's acting as this like hyperanalytical tool for tissue classification.

Exactly. It's doing math on the image.

But once an AI masters those straightforward diagnosis, the real test becomes predicting the future. Right. This is the holy grail Veramis discusses in the paper prognostication.

That's the frontier,

right? So the models are being trained to evaluate head and neck squis cell carcinoma or hnscc and basically predict its clinical behavior

right like will this specific tumor be aggressive

exactly will it metastasize

and predicting the clinical behavior or predicting the malignant transformation in oral premalignant disorders that requires an entirely different level of computational inference

because it's not just what is this it's what will this become

right I mean human pathologists are incredibly skil build at grading a tumor based on visual criteria. Right? We've been doing it a long time.

But an AI can detect subvisual patterns, things we literally cannot see. It can calculate the precise micrometer distance between every single tumor infiltrating lymphosy and the invading edge of the carcinoma.

Wait, really? Every single one.

Every single one. Yeah. Establishing this mathematical signature of the immune response that a human eye simply cannot quantify consistently.

That is mindblowing.

The goal is to look at a pre premalignant lesion and have the algorithm calculate the statistical probability of it turning into full-blown cancer just based on those microscopic spatial geometries.

You know, the way I think about the current state of AI in this field, it's like having a brilliant but highly specialized medical resident in your lab.

Oh, I like that comparison.

Right. Like for the textbook cases, this resident is lightning fast and flawless. They've memorized every pattern. But when it comes to predicting the nuance long-term behavior of a really tricky primol lesion that resident is still observing. They just don't quite have the decades of intuitive predictive wisdom yet.

That's spot on. The pattern recognition is dialed in. Absolutely. But predictive modeling is still highly uncertain.

Right.

Vermosis points out that while the models may soon benefit the prognostication of these malignant transformations, we are still figuring out how to make those models robust enough for clinical reliance.

It's a work in progress.

Exactly. We are building the architecture for prediction. But the foundation it sits on is currently highly unstable.

Which honestly brings us to the core conflict of our discussion today.

The big roadblock.

Yeah. Because if this algorithic resident is already this capable, if it can map the spatial geometry of lymphosytes and match human experts on complex classifications, why isn't it installed on the workstation of every single oral pathologist in the world today?

That's the billion dollar question,

right? You'd assume the bottleneck is computational, like we need faster GPUs, we need more memory, better silicon.

Sure, that's what everyone assumes.

But the roadblock standing in the way of this whole digital revolution is entirely analog. It's a physical laboratory problem, and it's known as pre-analytical variance.

And to really understand the magnitude of this problem, we have to look at the physical journey of a biopsy.

Walk us through it.

Turning a piece of oral tissue into a digital pathology image is a highly manual multi-step process. First, the tissue has to be preserved in formal Right.

Then it's embedded in paraffin wax,

sliced unfathomably thin on a micro, mounted onto a glass slide, and then subjected to H& staining.

Hemattoxylin and eosin.

Exactly. Hemattoxylin binds to the nucleic acids, turning the nuclei blue or purple, and eosin binds to the proteins, turning the cytoplasm pink.

Okay, pretty standard stuff for a lab.

It is. But here's a catch. Every single one of those steps introduces physical variability.

Okay, here's where it gets really interesting and I actually have to push back on the severity of this issue a little bit. Let's hear it.

Let's talk about that H& stain.

If you hand a human pathologist a slide that's like a little overstained, maybe the lab technician left in the EOS too long and the whole slide is just screaming hot pink today.

Happens all the time,

right? But the pathologist doesn't throw their hands up in defeat, do they?

No, of course not.

Their visual cortex just sort of white balances the image automatically. They look past the bad stain job and immediately identify the underlying square. a cell carcinoma.

So my question is, shouldn't a state-of-the-art hyper intelligent neural network be able to just look past simple differences in laboratory slide preparation just as easily as a human brain?

Well, what's fascinating here is that the assumption that AI sees the way we see is the fundamental trap in digital pathology.

Why? So

the neural network cannot just look past a heavy stain because the AI isn't actually looking at a picture of a cell.

Wait, what is it looking at then.

Okay. When a human looks at a slide, our brains process abstract concepts, right? We identify the shape of a nucleus or the structure of a keratin pearl regardless of its exact color shade.

Yeah. A circle is a circle whether it's light pink or dark pink.

Exactly. But a computer vision algorithm is processing a massive matrix of numerical data. It is looking at a grid of millions of pixels and every single pixel is defined by specific numeric values across the red, green, and blue channels.

Ah. So, it's not a picture at all. It's basically a giant spreadsheet of color codes.

Yes. Consider the math of it. If a human sees a slightly darker pink cytoplasm, it's the exact same conceptual image, just a bit darker.

Mhm.

But to an algorithm, if the eosin stain is left on for an extra 20 seconds, the red channel value of that pixel might shift from say 150 to 210.

Oh wow.

And when you multiply that mathematical shift across millions of pixels, the underlying data signature changes completely.

So, it's a totally different file to the computer.

The algorithm doesn't see a darker cell. It sees a fundamentally different mathematical environment.

Okay, I think I get it. It's like it's like teaching a state-of-the-art autonomous self-driving car to navigate flawlessly, but you train it exclusively on the roads of the Mojave Desert.

Oh, that's a great analogy,

right? It learns the visual signature of clear skies, dry asphalt, maybe some cacti, and it's perfect at it.

And then you take that exact same car with the exact same brilliant al algorithm and you drop it into a white out blizzard in Minnesota,

it's going to fail.

Exactly. The cameras and sensors are still working perfectly, but the environmental data the algorithm relies on like the contrast of the lane markers or the shade of the road, it's entirely obscured by snow.

Right?

The underlying data structure has completely changed. So, the car just crashes.

That captures the dynamic perfectly.

The algorithm hasn't lost its intelligence. It has just lost its statistical reference points. Does the input change too much?

Exactly. Because of the massive degree of pre-analytical variance, different formal and fixation times, varying section thickness, different brands of chemical stains, even

all those little analog choices,

right? The resulting whole slide images produced by different laboratories have massive data variability. A slide prepped in a clinic in New York looks mathematically foreign to an AI trained on slides from a university hospital in London,

which completely sabotages our ability to even measure how good these AI tools are, doesn't it?

It really does

because Vimis emphasizes that this variance makes comparing the performance of different artificial intelligence strategies incredibly difficult.

It's almost impossible,

right? Like if researcher A publishes a paper claiming their convolutional neural network detects oral cancer with 99% accuracy and researcher B claims their entirely different model hits 85% accuracy, we can't actually declare researcher A the winner.

Yeah. because they aren't running on the same track.

Exactly.

Researcher A likely trained and tested their algorithm on a highly curated localized data set from their own institutions laboratory

where the staining protocols are identical every single day.

Right? The model learned the specific quirks of that one lab's H& stain. It memorized the Mojave Desert to use your analogy.

Ah, I see.

But if you feed researcher A's algorithm a slide from researcher B's laboratory, the accuracy might plummet from 99% to 40%.

Just because the shade of pink is off by a few RGB values.

Exactly. The baseline data is so fractured that performance metrics become isolated anecdotes rather than, you know, universal truths.

And this analog variance causes an immediate massive digital consequence. It creates what the industry calls a data drought.

A huge problem right now.

Yeah. Now, if you're listening to this and you're running an independent pathology lab right now, you're probably thinking to yourself, "My specific staining protocols have worked perfectly for my team for 30 years.

Why change what isn't broken?

Exactly. My pathologists are happy. My patients are getting accurate diagnosis. Why on earth should I care if my slides confuse a piece of software?

It's a valid question from their perspective.

It is. But here is the uncomfortable truth. You might have thousands of slides in your archive, but your data is entirely siloed.

We don't just lack digital images. We lack robust, publicly available, standardized data sets.

And that's the key.

Building a model that can survive the blizzard requires exposing it to every conceivable weather condition. To train an open- source AI model to genuinely generalize, to teach it to mathematically ignore the endless variations of ancient e staining, tissue folds, and scanner artifacts.

Yeah,

you need to feed it millions of whole slide images.

Millions.

Millions. And those images need to come from thousands of different laboratories representing every possible variation of pre-analytic tissue preparation.

That's a staggering amount of data

and it gets harder. Furthermore, every single one of those millions of images must be flawlessly annotated by expert human pathologists to establish the ground truth for the algorithm.

Okay, we are talking about a logistical undertaking of just massive proportions here.

That's monumental.

And in the feeble oral and maxul facial pathology, those massive multi-institutional standardized public data sets simply do not exist yet.

Not even close.

The data is the actual bottleneck. You know, we constantly hear this narrative that the evolution of AI and medicine is waiting on the tech industry,

waiting for better chips. Yeah.

Right. Waiting for better graphics processing units, waiting for quantum computing, bigger server farms. But this analysis proves that the holdup isn't computational horsepower at all.

No, the transition from a highly controlled academic research setting into the chaotic real life environment of everyday clinical diagnostics that is tethered directly to the avail ility of varied ground truth data.

The data is everything.

You can possess the most sophisticated supercomputing cluster on the planet. But if the training data you feed into it is a fractured, unstandardized mess of idiosyncratic lab protocols, the predictive output will be entirely unreliable.

Garbage in, garbage out, basically.

Exactly. Addressing these pre-analytical concerns is not a secondary goal. It is the absolute mandatory prerequisite for bringing these AI models into the clinical environment.

The digital revolution has to start with and analog standardization.

It has to.

So what does this all mean? Let's distill the core insights from Dr. Vermus' work for our trailblazers listening.

Good idea.

The horizon of digital pathology is incredibly promising. We are looking in a near future where AI algorithms don't just match human experts on detecting squamous cell carcinoma, but actively map the subvisual architecture of a tumor micro environment.

All to predict whether a premolignant lesion will eventually threaten a patient's life. which is incredible.

It is. But that highly anticipated AIdriven future cannot be brute forced with better code. It relies entirely on standardizing the dusty, highly physical laboratory processes first. Right?

We literally cannot unleash the algorithm until we unify the physical data it consumes.

And that reality forces us to confront a pretty significant paradigm shift for the field.

It is

well if these computer vision algorithms require perfectly harmonized physical inputs to function at their maximum imum diagnostic potential. What is the ultimate fate of the independent localized pathology lab?

Oh wow, that's a dark thought,

right? Will the inevitable adoption of AI diagnostic tools force every laboratory in the world to abandon their unique traditional tissue preparation methods?

Methods they might have honed over decades.

Exactly. Will they have to drop all that in favor of a single rigid universallymandated global standard for tissue fixation and staining?

It's an incredible tension to consider. We might literally have to sacrifice the localized art of tissue preparation at the altar of algorithmic precision.

We might.

If we want the digital eye to see clearly, we may all be forced to start painting with the exact same chemical colors.

That's a great way to put it.

Keep questioning the process, trailblazers. Keep exploring that complicated intersection of analog medicine and digital technology. We'll catch you on the next deep dive on the Digital Pathology Podcast.