212: Digital Twins in Neuro-Oncology: A Systematic Review

Paper Discussed in this Episode: Digital Twins in Neuro-Oncology: A Systematic Review of Current Implementations, Technical Strategies, and Clinical Applications. Annie Singh, Fatima Ahmad Qureshy, Angelica Kurtz, Moinak Bhattacharya, Prateek Prasanna, and Gagandeep Singh. Radiology: Imaging Cancer 2026; 8(2).

Episode Summary: In this journal club deep dive, we explore a groundbreaking 2026 systematic review of digital twins in neuro-oncology. We step past the buzzwords and examine how exact virtual copies of patient brains are being built to safely simulate dangerous radiation regimens and drug combinations for highly aggressive tumors. This forces us to ask an uncomfortable question: Are we just slapping the label "digital twin" on static algorithms, or are we actually building living, continuously updating virtual copies of patient tumors? Furthermore, what happens to clinical ethics when a perfect simulation predicts a patient's tumor will resist every standard line of therapy before they even try it?

In This Episode, We Cover:

• Defining the True Twin: We break down what separates a standard, static computational model from a true digital twin. A real digital twin requires closed-loop optimization with continuous, real-time feedback from a patient's actual biological response—a critical feature shockingly missing in 13 out of the 21 reviewed models.

• The Dominance of Old-School Math: Why the most advanced simulations aren't relying solely on modern machine learning, but rather mechanistic models built on reaction-diffusion differential equations. We explain how these models calculate variables like tumor cell density, proliferation rate, and tissue carrying capacity to simulate literal physical pressure in the brain. Transparency and trust trump "black box" AI when neuro-oncologists are making life-altering surgical decisions.

• The AI Visual Forecaster: How cutting-edge AI diffusion models, like BrainMRDiff and ImmunoDiff, serve as hybrid partners to these math equations. These tools take complex biological calculations and generate high-fidelity, anatomically consistent visual MRIs to accurately forecast how a tumor will morph post-treatment.

• Grading Their Own Homework: A look at the PROBAST risk of bias assessment, which revealed that while outcome accuracy seems high, many models suffer from overfitting, data leakage, and a massive lack of external validation.

• The Big Bottlenecks - Broken Pipes and Locked Safes: We discuss the roadblocks keeping this out of the bedside. Specifically, the glaring lack of open-source code (only 6 of 21 studies shared theirs) makes standardization impossible. We also examine the engineering nightmare of multimodal data fusion—combining asynchronous streams of MRIs, genomics, and tissue pathology into a real-time model.

Key Takeaway: While digital twins represent a monumental leap toward true precision medicine, the field is currently bottlenecked by proprietary secrecy and broken data infrastructure. Until the scientific community embraces open-source code sharing and hospital systems solve the complex engineering challenge of real-time multimodal data integration, these revolutionary tools will remain isolated research projects rather than the living clinical tools they are meant to be

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

Transcript

Imagine for a minute, you know, that you could test a brutal radiation regimen on a patient

or uh even a highly toxic chemotherapy drug,

right? Exactly. But you do it on an exact virtual copy of a patient's brain. So you run the simulation, you watch how the tumor reacts, you adjust the dosage, and uh maybe you test a completely different drug combination. And you do all of this before you ever give the actual human patient sitting in the exam room a single drop of medicine.

It really sounds like science fiction, but it's I mean, it's happening.

It is. Welcome to the digital pathology podcast. Hello to all you trailblazers out there listening. Today, our mission is a journal club style deep dive into a massive leap forward in precision medicine. We are breaking down a brand new systematic review. It was published in March 2026 in the journal Radiology Imaging Cancer.

Yeah. And it's a huge paper.

It really is. It's titled Digital Twins in neurooncology, a systematic review of current implementations, technical strategies, and clinical applications and you know credit where credit is due. The authors here Annie Singh, Fatima Ahmad Koshi, Angelica Curtz, Moak Baracharia, Pratik Prasana and Gageep Singh. They really did some incredible heavy lifting.

Oh, absolutely. I mean they started by screening uh 73 articles from major databases and they had to filter out all the animal models, the basic science papers and you know those purely conceptual pieces that just didn't have real world clinical applications yet,

right? Getting rid of the fluff.

Exactly. So, they narrowed it all down to 21 studies that met their really strict criteria for patientspecific computational modeling in neurooncology,

which gives us a very clear mission for today's discussion. We are taking those final 21 studies to see how these virtual copies are actually being built, what they're actively predicting in hospitals right now, and crucially, what roadblocks are keeping this from becoming the standard of care for every single patient.

Because the term digital twin is, well, it's one of those buzzwords that just gets thrown around. constantly right now.

Oh, everywhere. You hear it in aerospace, you hear it in manufacturing, and now obviously it's making its way into medicine. But we really need to establish what that actually means in the context of brain cancer because I suspect it's uh it's a bit more than just a 3D rendering on a computer screen.

It is significantly more complex than a 3D rendering. And uh making that distinction is really fundamental here because the term has gotten so heavily diluted in popular media.

So what makes it a true twin then?

Well, based on the work flow blueprints that the authors outlined in this review. A true digital twin in neurooncology requires this massive fusion of patient specific data. It demands genomic sequencing, hystopathology from tissue biopsies, comprehensive patient history and it heavily heavily relies on imaging.

Really like MRI data.

Yeah, exactly. In fact, 17 of the 21 studies reviewed relied primarily on MRI data. Plus, they combine all that individual patient data with population level data to feed into computational model.

Okay, that reliance on MRI totally makes sense given we're talking about central nervous system tumors. You can't exactly monitor a brain tumor by just, you know, physically looking at it. You need that constant non-invasive visual data,

right? You need to see inside the skull.

But taking an MRI and making a computer model out of it isn't exactly a new concept. So, where does the actual twin aspect come in?

That is the million-dollar question. The defining characteristic, the thing that separates a regular computational model from a true digital twin is closed loop optimization.

Okay, let's unpack this. Closed loop optimization. Does that mean the model is like actually adapting over time?

Yes, that's exactly it.

So, if a patient's real world blood test comes back on a Tuesday with unexpected results or say their follow-up MRI shows the tumor didn't shrink as much as we expected, the model actually updates itself.

That is the core mechanism right there because a standard computational model is basically static. You feed it data on day one. It spits out a prediction and then its job is done.

It just sits there,

right? But a digital twin operates on a continuous feedback loop. The model makes a prediction, the clinician applies a treatment, and the patient's actual realworld biological response is fed right back into the model. The twin recalibrates in real time.

Wow.

Yeah. It's designed to be a living dynamic system for continuous clinical decision support. And that's why 12 of the studies specifically focused on highly aggressive tumors like glyopblastoomas. and high-grade glomomas.

Listening to you explain that feedback loop, it sounds exactly like a hyper advanced flight simulator, but instead of, you know, a generic Boeing simulator, it's built specifically for one pilot flying through the unique geography of one specific patient's brain.

That is a highly accurate way to visualize it.

And the closed loop part means if a storm suddenly rolls in during the real flight, it instantly starts raining in the simulator. You want to crash the simulator to find the safe route rather than risking the actual plane.

Exactly. You want the digital twin to absorb the failure of an ineffective therapy. That way, the patient doesn't have to endure the devastating side effects of a drug that was honestly never going to work anyway,

which is so critical for something like a glyopblastoma because those are just notoriously unpredictable.

A standard one-sizefits-all treatment protocol often fails because those specific tumors mutate and adapt so quickly,

right? They are incredibly heterogeneous. I mean, the genetic makeup of a glyopblasto can differ wildly Not just from patient to patient, but literally from one side of the tumor to the other.

Oh wow.

Yeah. So the need for a highly personalized continuously updating simulation is just most desperate in those specific high-grade arenas.

So if the stakes are that high and this simulator needs to be that incredibly precise, what is the actual engine powering it? Because I mean we are living in a moment where artificial intelligence and machine learning are dominating basically every single medical headline.

Oh for sure. AI is everywhere.

Right. So I would naturally assume these researchers are just feeding massive amounts of MRI data into some supercomput and letting a neural network figure it out.

You would assume that, but it was actually one of the most surprising findings in the entire review. The authors found that mechanistic modeling frameworks absolutely dominated the field.

Wait, really? Mechanistic models?

Yeah. 20 out of the 21 studies relied on mechanistic or biohysical frameworks, meaning they used old school physics and advanced mathematics rather than modern and machine learning.

Wait, AI is taking over everything else in healthcare right now. Why are old school physics and math equations beating out machine learning in something as cutting edge as digital twins?

What's fascinating here is that mechanistic models are deeply deeply grounded in known biological rules like how fluids move or how cells divide. The most common approach which was used in 15 of the study by the way relied on reaction diffusion models using partial or ordinary differential equations.

Okay, so it's heavily math-based

very much so. Rather than asking an AI to, you know, just find a statistical pattern in a million images, these researchers are explicitly programming the computer with the actual laws of human biology.

Give me an example of what that math actually looks like in practice because I'm trying to picture how an equation simulates a living tumor.

Well, I love this stuff, so I'll try not to geek out too much, but they use specific mathematical variables to represent physical biological realities. For instance, the equations use a variable P to represent the proliferation rate.

Okay. So that's like how fast the tumor cells are dividing.

Exactly. And then they use C of X comma T to represent the tumor cell density at a specific three-dimensional coordinate. That's the X and at a specific time which is T.

Got it. So it's tracking exactly how crowded a microscopic piece of brain tissue is at any given minute.

Yes. And that brings in another crucial variable which is K. K represents the tissue carrying capacity.

Carrying capacity. So think of K as like the maximum occup occupancy of a room.

That's a perfect way to put it. The surrounding healthy brain tissue can only provide so much physical space, right? And only so many blood vessels to support new cells. So when the tumor cell density approaches that carrying capacity, the cells have to push outward, creating literal physical pressure.

Wow. Okay.

And the differential equations calculate that pressure and predict exactly how the tumor will expand into the neighboring healthy tissue.

It's essentially treating the brain like a complex fluid dynamics and structure. engineering problem.

Yeah,

that is fascinating. But I have to ask, why do clinicians prefer this over an AI that might be able to, you know, crunch the data a lot faster?

Two words, transparency and trust. A neurooncologist can look at a reaction diffusion equation and understand the exact biological rationale behind the prediction.

Whereas AI is just a mystery,

right? AI is famously a black box. It might give you a highly accurate prediction, but it cannot explain its underlying logic to you. When a surgeon is deciding whether to remove a completely healthy looking piece of brain tissue because a computer predicts the tumor will spread there in 3 months. They need to trust the logic.

Yeah. You cannot base a permanent lifealtering surgical decision on an algorithm that basically just says, "Trust me, bro."

Exactly. You need the receipts.

So, if that transparency is so vital to doctors, where does AI even fit into this? Did the review find any role for machine learning or is it just completely sidelined in neurooncology?

Oh, it's definitely not sidelined. It's just being used differently. Six of the studies incorporated AI mostly as hybrid models working alongside the math we just talked about. But the review actually highlights a very exciting, very significant shift toward AIdriven diffusion models for image generation.

Image generation.

Yeah. The authors specifically single out newer models like brain miff and amodiff.

Wait, when you say diffusion models for image generation, are we talking about the same underlying technology people use to generate like weird AI art on the internet but applied to MRI? The underlying architecture is actually very similar. Yeah. But obviously the application is vastly more rigorous. These AI models are generating highly accurate anatomically consistent pre-treatment and post- treatment imaging. They are literally providing a visual forecast of the tumor's response for the doctor to see.

But why do they need a totally new model like Brain Mr. DIFF for that?

Couldn't older AI models just generate a picture of a shrinking tumor?

Well, older baseline models often struggle with the rigid anatomical reality of the brain. pain. When an older AI tries to predict what an MRI will look like after, say, 3 months of chemo, it might accidentally blur the boundaries of the skull.

That's not good.

No. Or it might hallucinate changes to the natural folds of the healthy brain tissue. It lacks anatomical fidelity. Brain MRF solves this problem. It maintains the exact physical geometry of the patients specific brain structures while solely predicting the morphological changes of the tumor itself.

So you have the mechanistic models doing the transparent rig ous biological math to calculate the cellular pressure and the growth and then the AI models take those calculations and generate a highfidelity visual MRI for the doctor to actually look at. That is a brilliant division of labor.

It's a highly effective hybrid approach. The paper even includes this incredible visual example showing a simulation where they test multiple therapies. You see the patients pre-treatment MRI and then the AI generates predicted post treatment MRIs for treatment one, treatment two, and treatment three.

That's Wild.

Yeah. And you can visually see that treatment three most effectively reduces the tumor burden and the patient didn't have to endure the trial and error of the first two options. Amunif does something similar for non small cell lung cancer visually differentiating between patients who will respond to amunotherapy and those who just won't.

Okay. So if we trust this math and we trust this visual generation, what are researchers actually asking these twins to calculate on the clinical front lines today? Like what are the use cases?

The review categorizes the clinical applications quite clearly. The absolute most common application which they found in eight of the studies is tumor growth simulation.

Makes sense.

Yeah. This is essentially answering the question if we delay surgery for 3 weeks to let the patient stabilize exactly how much will this mask grow and in what specific direction?

Closely followed by treatment response is

correct. Seven studies focused on treatment response. That's simulating how a specific tumor will react to radiation immune system. interactions or tracking how effectively a drug will actually transport into the tumor.

It feels very similar to hyperlocal weather forecasting.

Simulating tumor growth is like predicting the exact path and intensity of a hurricane over the next 5 days.

That's a great analogy.

And simulating treatment response is predicting what happens if we now attempt to seed the clouds to disrupt the hurricane. Will it dissipate or will it just fracture and cause damage somewhere else?

That translates perfectly. And

meaning the tumor isn't just taking up physical real estate in the skull. It's actively shortcircuiting the brain's electrical wiring.

Precisely. Think of the brain's neural network like a city's power grid. A growing tumor acts like a localized blackout. It forces electrical signals and cognitive traffic to reroute. The digital twin can predict not just the physical growth of the tumor, but the functional cognitive collateral damage it will cause to the rest of the network.

Predicting cognitive collateral damage before it happens. I mean, that is next level precision. medicine. But we have to talk about accuracy. When you build a model this complex, the margin for error must be massive. Did the authors evaluate how reliable these models actually are?

They did, and they used a very specific standardized framework to do it called the Probast.

Probast.

Yeah. It stands for the prediction model risk of bias assessment tool. It evaluates four domains: participants, predictors, outcome, and analysis. It's essentially designed to uncover whether a predictive model is actually reliable or if the researcher's unknown ly skewed their own results.

So looking at the Probast report card for these 21 studies, what did the authors find? Did they pass?

Well, the results reveal a pretty significant divide. On the positive side, the models show excellent predictive accuracy regarding the outcomes. When evaluating the predictors and the outcome domains, the risk of bias was generally low. The variables they are tracking, like that cellular proliferation we talked about, they correlate highly with actual clinical data.

That sounds incredibly promising but uh I hear a however coming.

You do. However, there were major systemic red flags regarding the participants and the analysis domains. For the analysis domain specifically, nearly a quarter of the studies had a high risk of bias and a large chunk were simply categorized as unclear.

Wait, if the math is so solid and the outcomes are so accurate, how do they fail the analysis portion? Are they mishandling their own data?

And in many cases, yes. A really common issue leading to a high risk of bias and analysis is over overfitting or data leakage.

Overfitting. What does that look like in this context?

Imagine a research team builds a digital twin model and they train it using the MRI data of 10 specific patients. The model learns those 10 tumors perfectly. But if the researchers then test the model's accuracy by asking it to predict the outcomes of those exact same 10 patients.

Oh, I see. They're essentially letting the model grade its own homework.

Exactly.

It hasn't actually learned the universal biological rules of brain. cancer. It just memorized the past history of those 10 specific people. So if you apply that same model to an 11th patient who just walked into the clinic today, it might completely fall apart.

That is exactly the problem. And it highlights a massive lack of external validation in the field right now. A digital twin might work flawlessly on a tightly controlled homogeneous data set in a research lab, but fail completely when applied to a diverse patient population in a chaotic realworld hospital environment.

Which naturally leads to the biggest question of this ent entire review. If these digital twin models possess this incredible predictive power, why aren't neurooncologists using them for every single patient right now? What are the actual bottlenecks keeping this out of the bedside?

Annie Singh and her co-authors were very explicit about the limitations holding the field back. The single biggest hurdle is a shocking lack of reproducibility and transparency in the software itself. Out of the 21 studies included in this review, only six provided publicly available code.

Wait. Only six out of 21.

Yep, only six.

If the code is locked away, how is the scientific community supposed to verify the math? You can't test for that data leakage or overfitting if you can't even see the algorithm.

You literally can't. This method heterogeneity where every lab is building custom proprietary models and keeping their methods a secret. It makes standardization completely impossible. If you don't share the code, other hospitals cannot test your digital twin on their own patient cohorts to see if it actually holds up. But there is a secondary bottleneck that is arguably even more critical and it goes all the way back to our very first definition of what a digital twin is supposed to be.

The closed loop optimization.

Yeah.

The ability to update in real time based on new patient data.

Exactly. Out of the 21 studies published under the label of digital twins, only eight actually reported closed loop calibration.

Are you kidding me? So, the vast majority of these so-called digital twins are missing the very feature that makes them a twin.

It's true.

Are we just slapping the buzz? label of digital twin onto things that are really just static one-off algorithms because if it doesn't have that real-time closed loop feedback, isn't it technically just a standard model?

I completely validate that concern. It is a profound structural limitation and honestly it stems from a massive engineering challenge which is multimodal data fusion. It is incredibly difficult to take disperate types of data imaging, genomics, tissue pathology and fuse them together smoothly in real time.

I can totally see That's a nightmare in a real hospital setting. The timelines simply don't match up. The MRI gets uploaded to the electronic health record on a Tuesday. The pathology report from the biopsy might trickle in on say Thursday. And then the deep genomic sequencing of the tumor might take 3 weeks to come back from some external lab,

right? And currently the digital infrastructure to feed all of those mismatched asynchronous data streams automatically and continuously into a unified computational model. It just doesn't not exist in most clinical settings.

It's just broken pipes everywhere.

Exactly. The review discusses future solutions like edge computing and cloud-based frameworks to handle the processing load, but currently the pipes simply aren't connected. Encilico trials are brilliant in a vacuum. But until the field embraces open- source code sharing and solves the engineering nightmare of real-time data integration, these tools will remain isolated research projects rather than the living clinical tools they're meant to be.

That is a sobering reality check, but it also lays out a very clear road map for the future of neurooncology. Let's recap the journey we've taken today. Digital twins represent a monumental leap toward true precision medicine. They are currently powered by highly trusted mathematical mechanistic models that simulate actual physical and biological pressures. While newer AI diffusion models like BrainMR diff are stepping in to generate highfidelity visual forecasts.

Yeah. And we've seen that while they excel at simulating tumor growth and mapping out fluid dynamics, the field is frankly currently bottlenecked. To move from the lab to the clinic, researchers must mandate open- source code sharing to allow for external validation. And hospital systems must fundamentally restructure how they handle multimodal data.

Which brings us to a crucial call to action for you, the trailblazers listening right now as you navigate this evolving landscape. Think about the systems you interact with daily in your own specialties. The next great frontier in healthcare isn't just about building smarter algorithms. It's about standardization. How do we build an infrastructure where a digitized pathology slide, a genomic sequence, and an MRI scan can instantly and seamlessly talk to the exact same digital twin.

Solving that data integration problem really is the key to unlocking this technology.

It absolutely is.

And as we push toward a future where these perfect simulations do become a reality, I want to leave you with a challenging ethical dilemma to ponder. Let's assume we solve the data fusion problem. We build the perfect digital infrastructure and we create a flawlessly accurate real-time digital twin for a patient with an aggressive glyopblasto.

Okay, I'm with you.

What happens to the clinical decision-making process when this highly accurate twin predicts with near absolute certainty that the tumor will resist every single standard line of therapy?

Wow, that changes the entire conversation in the exam room.

It really does. How does that forecast alter the ethical landscape of care? Does the twins prediction justify fasttracking a patient immediately into highly experimental potential dangerous treatments? Or does leaning entirely on the simulation run the risk of creating a self-fulfilling prophecy, prematurely limiting a patient's hope and options before they've even had a chance to try the established standard of care?

That is a heavy, profound question regarding the weight of knowing the future and a perfect place to wrap up our analysis. Thank you all so much for joining us for this journal club deep dive into the virtual front lines of neurooncology. Keep pushing the boundaries, keep questioning the systems around you, and we will catch you on the next one.

Keep question the data and keep learning.