Digital Pathology Podcast
Digital Pathology Podcast
187: AI vs. Human Pathologists: Who Sees the Biology of Glioblastoma Better?
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Paper Discussed in this AI Journal Club:
Episode Summary: In this journal club deep dive, we step out of the theoretical future of AI and look at a direct, hard-data showdown between artificial intelligence and the human eye. We examine a groundbreaking 2026 study on Glioblastoma Multiforme (GBM) that forces us to ask an uncomfortable question: What happens when the AI and the human completely disagree? And more importantly, is it possible that the AI is detecting a biological reality that experienced human pathologists are entirely missing?
In This Episode, We Cover:
• The "Boss Battle" of Neuro-Oncology: Understanding Glioblastoma Multiforme (GBM), the most aggressive primary brain tumor in adults, and why precise prognosis dictates the entire treatment strategy.
• The Big Three Biomarkers (The Speedometer, The Brakes, and The Neon Sign):
◦ Ki67: The "speedometer" of the tumor, marking active cell proliferation.
◦ p53: The "guardian of the genome," acting as the emergency brakes for damaged cells. In GBM, these brakes are often broken or mutated.
◦ PHH3: A specific "neon mitosis tracker" that lights up dividing cells, offering a cleaner alternative to traditional manual counting.
• The Showdown - Humans vs. AI: Two experienced pathologists go head-to-head with an AI digital image analysis system (QuantCentre module by 3DHISTECH) on 20 adult GBM cases, looking at both 1 mm² and 7 mm² tumor hotspots.
• Round 1 - The Shocking Lack of Concordance: The AI and human pathologists had practically zero statistical agreement (Cohen's Kappa) on the raw numbers. The human eye acts interpretively, filtering out background noise, while the AI calculates literal pixel intensity.
• Round 2 - The AI's "Aha!" Moment: Biologically, a high proliferation rate (Ki67) must correlate with high mitosis (PHH3). Human pathologists failed to find any statistically significant link between these markers. The AI, however, found strong, biologically accurate correlations between Ki67 and PHH3, and between PHH3 and p53.
• The Future of the Lab: Why AI shouldn't replace pathologists, but rather serve as a hyper-sensitive tool to uncover hidden data patterns and personalize medicine. We also discuss the major roadblock preventing immediate clinical rollout: color standardization and image quality.
Key Takeaway: The lack of agreement between humans and machines doesn't mean the AI is wrong. By successfully identifying crucial biological relationships that humans missed due to attentional fatigue and subjectivity, the AI proved its data might actually be closer to the biological truth than our current gold standard.
Question of the Week for Our Trailblazers: Should we stop asking if the AI is as good as the human, and start asking if the human is actually precise enough to judge the AI? Let us know your thoughts!
Hello and welcome back to the digital pathology podcast. Uh to all the trailblazers tuning in today, whether you're at the lab bench, maybe the reading room, or even the developers desk, we are really glad to have you with us.
Yeah, it's always a pleasure to be here with this community. We have um quite a session lined up today, a proper journal club deep dive.
We really do. You know, if you're here, it's because you care about where healthc care is going, not just where it's been.
And we spend a lot of time in this industry talking about the concept of AI.
Uh the potential, the whatifs,
right? The hypothetical future.
Exactly. We talk about it like it's this magical future state. But today, we're grounding that. We are looking at a direct head-to-head showdown.
We are. This isn't a theoretical discussion today. This is about hard data. Specifically, we're looking at the intersection of high stakes oncology, glioblastoma specifically, and the emerging dominance of digital image analysis. And the central tension we're exploring is actually surprisingly gritty because usually uh Uh when we see these studies, the headline is just, you know, AI matches human performance. Everyone claps, we publish, and we move on.
Yeah. The classic victory lab,
right? But the study we're covering today forces us to ask a much more uncomfortable question.
It does. It forces us to ask, what happens when the AI and the human don't agree? And more importantly, and this is the core of it, is it possible that the AI is seeing a biological reality that the experienced human eye is just completely missing?
Oh, I love that setup. Let's get right into the specific specifics. So we are analyzing a study titled artificial intelligence-based digital image analysis for assessing KI67 P-53 and PHH3 expression in glioblastoma multiform.
It's a mouthful but a very important paper and it was published just recently in February 2026.
Thought off the press. It's in the journal of the college of physicians and surgeons Pakistan uh volume 36 issue 2
and we absolutely need to give credit to the research team behind this work. This comes from Tuba Devim Gam Urkalink and Sier Seem Tonser and and they're operating out of the department of medical pathology at Sigly training and research hospital in Isizmir Bakirke University in Turkey.
It's a very solid team and they tackled a notoriously difficult subject here.
So, what was their mission? Because looking at the title, it sounds like they're just, you know, automating cell counting. But I have a feeling there's more to it than just speeding up a boring task.
Oh, there is much more. Their primary goal was to compare the gold standard, which for decades has been ual quantification by experienced pathologists directly against an AI based analysis system
to see if they match up,
right? To see if digital pathology offers a new truth or at least a different vantage point in the prognosis of aggressive brain tumors.
Okay, let's unpack the clinical context first for our trailblazers. We're dealing with glioblastoma multiform or GBM.
Correct. GBM is basically the boss battle of neuro-oncology. It's the most common and unfortunately the most aggressive primary brain tumor in adults. The mortality rate is very high and the prognosis is generally poor.
And nuances matter here with this cohort. They didn't just grab every brain tumor they could find for the study, did they?
No, they filtered it carefully, which is important for the scientific validity.
They specifically looked at adult cases. They excluded pediatric cases entirely to keep the biological variables consistent,
right?
And they focused on what we call Idh wild type cases.
Okay? So, when you have a tumor this aggressive, getting the prognosis right basically Knowing exactly how fast it's moving is everything for the treatment plan.
It dictates the entire strategy. Yes.
And to do that, pathologists look for specific biomarkers. So this study focused on the big three for GBM.
The big three. Exactly.
Can you walk us through these? I want to make sure we understand not just what they are, but why they actually matter to the patient.
Certainly. These are immunohistochemical markers or IHC markers. Let's start with K67. In the lab, we often just call this the proliferative index.
I've heard it des described as the speedometer of the tumor.
That is a perfect analogy. Kai67 is a protein that is strictly associated with cell proliferation. It's present during all the active phases of the cell cycle G1, S, G2, and mitosis,
but not when it's resting.
Crucially, no, it is absent in resting cells, the GOVAS. In GBM, you typically see expression levels between 15 and 40%.
So, generally speaking, the higher the Ki 67 level, the larger the tumor volume and the worse the prognosis. literally tells you how fast the car is driving.
Exactly.
Okay, so that's the speedometer. What is marker number two?
P-53. This is often called the guardian of the genome.
The guardian. That sounds very dramatic.
It is dramatic because its job is absolutely vital. The TP53 gene is a tumor suppressor. When it works correctly, it acts like an emergency break.
Meaning, if a cell is damaged,
if a cell has DNA damage, p-53 stops it from dividing, so it can either be repaired or destroyed.
But in glioblastoma, the guardian isn't doing its job. The guardian is usually asleep at the wheel or actively sabotaged. In about 84% of GBM patients, this p-53 pathway is deregulated.
Wow. 84%.
Yeah. Often you have a mutant form of p-53 that just accumulates in the nucleus. So instead of suppressing the tumor, you get this high expression of mutant protein that directly correlates with malignancy.
So we have a fast car with Ki 67 and broken brakes with P-53.
Exactly.
That brings us to the third one, PHH3. This one feels a bit more niche than the other two. To me,
it is more niche, but for the pathologists listening, this is a really cool marker. PH3 stands for phospholated histone H3.
And what's its job?
It's a highly specific mitosis tracker. Now, let's step back for a second. Traditionally, to see how fast a tumor is growing, a pathologist looks at a standard H& stained slide, the standard purple and pink one, and tries to count mitotic figures.
You literally just count the cells that look like they're active. splitting in half,
right? But have you ever tried to do that?
I have not.
It is notoriously difficult.
A cell undergoing apoptosis dying can look a lot like a cell undergoing mitosis or a cell that got crushed during the biopsy processing can look like mitosis.
Oh, I see.
So the interobserver concordance basically how often two pathologists agree on the count is historically very low. It's highly subjective.
So PH3 is essentially a cheat code for this
in a way. Yes. It's an antibody that specifically lights up histone proteins only when they are phosphorylated during mitosis. It's like putting a bright neon sign on the dividing cells.
So, it doesn't light up for dying cells.
Nope. Doesn't light up for resting cells either. It's supposed to be much cleaner and more specific than the human eye on a standard stain.
Got it. So, to recap, we have the speedometer, which is K67, the broken brakes, P-53, and the neon mitosis tracker, PH3.
Right. And the researchers wanted to see, can an AI count these as well as a human expert?
Precisely. Let's get to the show. The methodology. They had 20 GBM cases. Who were the contenders in the ring?
In Cornerblue, we had the humans, two experienced pathologists. They performed manual counting in a blinded manner,
meaning they didn't know the patient outcomes,
right? And they didn't know what the AI had scored. They selected hot spots, the areas where the staining looked the most intense, and they counted 100 tumor cells to calculate a percentage.
Very standard practice. And then corner red,
the AI They digitize the slides using a high-end scanner and analyze them using the quant center module. This is part of the the clinical viewer software by 3D heist.
Now, here's where it gets really interesting to me. They didn't just run the algorithm once. They messed around with the field size.
Yes, they did.
They looked at a small area, one square millimeter, and a much larger area, seven square millime. Why the two sizes?
That is a great question, and it addresses the massive headache of tumor heterogeneity.
Heterogeneity, which basically means The tumor is just a mess, right?
Scientifically speaking, yes. Glioblastomas are chaotic. One part of the tumor might look drastically different from another part just a few millimeters away.
So they wanted to see if zooming out helps.
Exactly. They wanted to know if we zoom out to seven square millime, does the score change compared to that tight 1 mm hot spot, the sample size matter for accuracy?
Okay, predictions time. If I'm a listener, I'm assuming the AI and the humans are going to be pretty close. Maybe off by a percent or two. I mean, a stained cell is a stained cell, right?
You would hope so. But this is where we get into round one, concordance, and the results were well, they were jarring.
Jarring is definitely one word for it. They basically didn't agree at all.
The statistical agreement was very low. They measured this using something called Cohen's Kappa.
For our trailblazers who aren't statisticians, give us the quick cheat sheet on kappa.
Sure, a kappa of 1.0 is perfect agreement. A kappa of zero is random chance. When they looked at Kai 67, the speedometer The agreement between the AI and the pathologists was.175.175. That is barely better than flipping a coin.
It is technically considered slight agreement. Statistically insignificant.
Wow. And the other markers
similar story for P-53 in the 1 mm area. The cappa was 104. Practically zero agreement.
Unbelievable. What about the neon sign PHH3?
PHH3 was a little better at 418, but still no significant concordance in the smaller area.
This feels like a crisis. Does this mean the AI failed or does it mean the humans failed?
It's not necessarily about failure on either side. It's about the fact that the methods of quantification are just fundamentally different. The source material points out that image quality and color standardization are major major hurdles here.
How so? Like how does the AI see it differently?
Think about how an AI sees. It looks at strict pixel intensity. It has a rigid threshold. If a pixel is darker than value X, it's positive. If it's lighter, it's negative. Now, think about the human eye.
Filter things out naturally.
Exactly. We do a massive amount of subconscious processing. We ignore background noise. We compensate for a weak stain.
Right.
A human might look at a cell and say, "That's technically brown, but it looks like a macrophage, not a tumor cell, so I'm ignoring it." The AI is literal. The human is interpretive.
So, they just aren't speaking the same language.
Not yet. But, and this is a silver lining, there was consistency. While the AI and humans didn't agree with each other, they completely agreed with themselves.
Meaning what exactly?
Meaning the AI's score for the small 1 millm area perfectly matched its own score for the big 7 mm area.
Oh, I see.
And the human score for the small area matched their score for the big area.
Ah, so carefully chosen hotspots actually do work.
Exactly. The study concludes that if you select your hot spot, well, if you really find the most active part of the tumor, you don't necessarily need to scan massive areas to get a representative score,
which is great news for workflow efficiency in the lab.
Huge news.
Okay, so round one goes too well. Nobody or maybe it's a draw. They are consistent within their own logic, but they do not agree with each other. But round two, round two is where the real story is.
Yes,
this is the aha moment of the entire deep dive today.
This is definitely the twist and for me this is the most crucial part of the findings.
They looked at the correlations between the markers. So logically, does a high Kai 67 score predict a high PH3 score?
Right? Let's look at the basic biology. Kai67 marks cells in the cell cycle. PHH3 marks cells in mitosis.
You cannot have mitosis without being in the cell cycle.
Exactly. They are inextricably linked processes. If the speedometer is high, the neon mitosis signs should absolutely also be flashing.
So when they analyze the human pathologist's data, did they see this link?
No.
Wait, really? So to the human eye, the speedometer and the mitosis tracker had nothing to do with each other?
According to the manual counts in the study, yes. They appeared to be entirely independent variables. There was no statistically significant correlation found by the humans.
But that doesn't make any sense. Biologically, that's impossible.
Correct. It completely defies the known kinetics of the cell cycle.
Okay. So, what did the AI find?
The AI found strong statistically significant correlations. Specifically, the AI showed a strong positive correlation between Kai 67 and PHH3. The R value was 750, which is quite high.
And it didn't stop there. Right.
No, it didn't. The AI also found a significant link between PHHH3 and P-53.
Okay, let's just pause and let that sink in for a second. The AI found a connection that makes absolute biological sense, but the highly trained humans missed it.
That is the big takeaway here. The AI's data aligned perfectly with the laws of biology. The human data did not.
Why are we just bad at counting?
We aren't machines. And that's exactly the point. The source explicitly states that traditional morphological mitotic labeling by the human eye may be prone to error
because we get tired.
We get tired. We have attentional fatigue
and subjectivity.
Huge subjectivity.
Uhhuh.
Maybe we overcount based on what we expect to see. Or maybe we miss subtle staining that the digital algorithm picks up effortlessly.
So the AI is just more sensitive.
It is likely much more sensitive. It's detecting a biological reality. The concrete link between proliferation and mitosis that was basically invisible to the manual review.
That is profound. It basically says says the AI's truth might be closer to the actual biological truth than our current gold standard.
It definitely implies that AI provides a more consistent data set. And when you look at the link between PHH3 and p-53 that the AI found, remember p-53 regulates the cell cycle,
right? The breaks.
If p-53 is mutated and accumulating, meaning high expression, the cell cycle goes unchecked. That leads directly to more mitosis. So high pH3. Again, the AI data beautifully supports this biological narrative.
It's validating the biology via the algorithm.
Exactly. It proves that the algorithm isn't just spitting out random numbers. It's detecting a real signal where we just saw noise.
This brings us to the implications for our trailblazers listening. What does this mean for the future of the lab? If I'm a pathologist listening to this right now, should I be worried or should I be excited?
You should be excited, but realistically cautious. The authors conclude that AI offers improving objectivity and individualization. It completely remove the interobserver variability
because if you run the same slide through the AI twice, you get the exact same number
every time. If you give it to two different pathologists, you might get very different numbers.
But they also mentioned a big roadblock in the paper. Standardization. We can't just roll this out to every clinic tomorrow, can we?
No, absolutely not. The study was very clear on this limitation. For AI to reach routine clinical use, we have to solve the image quality and color standardization issues.
Right? Because if the slide is stained slightly darker blue in one lab versus another across town, the AI might freak out.
It's the pre-analytical variables. Humans are incredibly good at compensating for a bad stain. We can look at a slide and say, "Oh, the technician left this in the hematoxilin a little too long." And we mentally adjust.
But the AI takes the pixel values literally.
Exactly.
Now, they tried to fix this with the H score in the study, didn't they?
They did. The H score or his score is a method where you multiply the percentage of positive cells by the intensity of the staining, usually graded 0, 1, 2, or 3.
So, it's supposed to weigh the intensity to balance it out. Did it help?
Not enough to bridge the gap. Unfortunately, even using this standardized scoring method, the AI and humans still diverge significantly.
That's fascinating.
It reinforces that the discrepancy isn't just about counting. It's fundamentally about how we perceive intensity versus how a computer measures it. Until we standardize the actual staining process itself across labs, the AI will struggle to be universal.
But if we can solve that, if we can get the color standardized, the future of prognosis looks very different,
very different. The authors note that recent advances in machine learning are already improving survival prediction by integrating this exact kind of data. We are rapidly moving toward individualized medicine.
What does that actually look like in practice for a patient?
Well, instead of just categorizing a tumor as grade four and giving a standard protocol, the AI might analyze the precise relationship between p-53 and pHh3 and say this specific tumor profile suggests resistance to standard chemotherapy.
It allows for the detection of patterns that are just completely invisible to manual review.
Exactly.
It's the hidden patterns. That's what excites me the most. We think we see everything through the microscope, but we're really just scratching the surface of the data.
And that connects back to the bigger picture of where healthcare is going. It's not about replacing the pathologist. Not at all. It's about giving the pathologist a tool that sees into a spectrum of data. we simply cannot access with our eyes and brains alone.
So to synthesize what we've learned today from Drim and colleagues, we took 20 GBM cases. We pitted humans against machines. They didn't agree on the raw numbers, but and this is the massive, but the AI successfully identified biological relationships, specifically the link between proliferation and mitosis that the humans completely missed.
That is the absolute key takeaway. The lack of concordance doesn't mean the AI is wrong. In this case, the Internal biological logic strongly suggests the AI might actually be more right.
It's a bit humbling, isn't it?
It is very humbling. But science should be humbling. If we aren't being challenged by our own tools, we aren't learning.
I love that. Before we sign off, we always like to leave our trailblazers with a thought to chew on. Something that takes the source material and pushes it just one step further.
I've actually been thinking about this correlation finding all week. You know, we spend so much time in this field validating AI against the human gold standard. We write endless papers asking did the AI get the same number as Dr. Smith,
right? That is always the benchmark for success.
But if the human data doesn't correlate with the basic undeniable biology of the cell cycle and the AI's data perfectly does, here is the provocative question.
Let's hear it.
Should we stop asking if the AI is as good as the human and start asking if the human is actually precise enough to judge the AI?
Oh, that is the question. Maybe the gold standard just needs to be melted down and recast
or perhaps just upgraded with some new sensors.
Listeners, we would love to know what you think about this. If you want to dig into the data yourself and we highly encourage you to check out the full paper again that is by DevM in the journal of the college of physicians and surgeons Pakistan.
The statistical breakdown in the methodology section is really worth a read if you enjoy the technical side of things.
Thank you so much for joining us for this session of the digital pathology podcast. Keep analyzing, keep questioning and keep Keep blazing those trails.
Until next time.
Bye for now.