When Heart Fat Becomes a Thermometer: AI, CT Scans, and the Future of Plaque Detection

Sam: Did you know that the fat wrapped around your heart might actually be a biosensor capable of predicting a heart attack long before a vessel gets blocked?

Maya: It is a fascinating concept, and it is exactly what we are breaking down today from a paper titled Pericoronary adipose tissue imaging: quantitative assessment, artificial intelligence integration, and therapeutic modulation, published in The British journal of radiology.

Sam: And before we dive into the details, a quick reminder that our voices are AI-generated, and this episode is a summary of that publicly available medical document.

Maya: To set the stage, we are talking about coronary artery disease, which remains a leading cause of global morbidity and mortality. When a patient comes into the clinic with chest pain, the first-line investigation is usually coronary computed tomography angiography, or CCTA.

Sam: Right, a routine heart CT scan. They inject some dye, take the pictures, and look for blocked pipes. If a major artery is blocked, they can go in and open it up.

Maya: Exactly. Modern multi-detector CT and photon-counting detector CT give us incredibly high spatial and temporal resolution to see significant luminal obstruction. But here is the massive clinical problem: a substantial number of acute cardiac events actually occur from the rupture or erosion of smaller, non-obstructive plaques.

Sam: So, people are having heart attacks from plaques that the standard interpretation of the scan might not flag as an immediate emergency, because the vessel isn't fully blocked.

Maya: Precisely. Clinicians do have ways to spot high-risk features on a scan, like low-attenuation plaque, positive remodelling, spotty calcification, and the napkin ring sign. But these are qualitative assessments. They are limited by modest sensitivity and positive predictive value.

Sam: They are essentially just looking at the shape and density of the plaque itself and making an educated guess. Which brings us to the core issue the paper addresses. It all comes down to inflammation.

Maya: Yes. Inflammation is the key pathological process driving the initiation, progression, and rupture of these atherosclerotic plaques. If we could just measure the activity of coronary inflammation, we could evaluate which vessels and plaques are vulnerable and at risk of rupturing.

Sam: Which creates the opportunity for early preventive treatment. But how do you measure inflammation non-invasively? That is where the fat comes in. Specifically, pericoronary adipose tissue.

Maya: Correct. We call it PCAT. Adipose tissue is not just inert storage. It is a metabolically active organ. PCAT is a biologically distinct sub-depot of epicardial adipose tissue. It has unique paracrine interactions with the vascular wall.

Sam: Let us define exactly where this fat is. The paper gives a very specific geographic boundary for what counts as PCAT.

Maya: It is defined as the adipose tissue directly surrounding the vascular wall within a perpendicular distance from the vessel equal to the luminal diameter of the vessel, up to a 2 cm threshold. For vessels with a luminal diameter greater than 2 cm, it is simply the fat within a 2 cm perpendicular distance.

Sam: So it is hugging the artery tightly. In the larger coronary arteries, it actually touches the outer layer, the adventitia. But in smaller vessels, the fat cells might actually be incorporated right into the vascular wall.

Maya: That tight physical relationship is what makes it a biosensor. In a healthy state, PCAT plays a highly favourable role. It releases adipokines, cytokines, and gaseous mediators that exert vasodilatory, anti-inflammatory, and anti-oxidative effects.

Sam: The paper uses a great phrase for this. It says PCAT participates in bidirectional outside-in and inside-out signalling to regulate vascular tone. So the artery talks to the fat, and the fat talks to the artery.

Maya: Exactly. But in a pathological state, when the vessel becomes diseased, that conversation turns toxic. The inflamed arterial wall releases inflammatory and oxidative mediators, like interleukin-6 and tumour necrosis factor-alpha, directly into the surrounding PCAT.

Sam: And how does the fat react to that inflammatory bath?

Maya: It alters the fat's phenotype. It impairs adipogenesis, promotes lipolysis, and increases water content. You get a gradient of smaller, lipid-depleted adipocytes right next to the inflamed vessel, which transition back to larger, normal adipocytes as you move radially outward.

Sam: So the fat closest to the sick artery shrinks, loses its lipid content, and fills up with water and inflammatory cells. And this structural gradient acts as a thermometer of vascular inflammation.

Maya: That is the perfect analogy. And the brilliant part is that this shift from lipid-rich to watery is detectable on the routine clinical CCTA scans we already do every day.

Sam: How exactly does the CT scan see this? We have to talk about Hounsfield Units, right? The scale we use to measure radiodensity on a CT.

Maya: Yes. Under physiological conditions, healthy PCAT has a high lipid and low aqueous content. This gives it low attenuation values on a CT scan, specifically around -190 to -80 Hounsfield Units.

Sam: Negative numbers because fat is less dense than water, which sits at zero.

Maya: Right. But during vascular inflammation, when the fat cells shrink and water and inflammatory cells infiltrate the tissue, the local density increases. The attenuation values shift toward the higher, less negative end of the range. Specifically, from -190 to -30 Hounsfield Units.

Sam: So if the fat around the artery measures closer to -30, it is inflamed and watery. If it measures closer to -80 or lower, it is healthy and lipid-rich. That sounds straightforward, but I imagine actually measuring this consistently across different hospitals is a nightmare.

Maya: It is a massive technical challenge, which is why raw attenuation values alone are not enough. They are too influenced by technical factors like tube voltage, scanner type, and reconstruction kernels, not to mention anatomical variations.

Sam: So how do they solve for all that noise? The paper introduces something called the Fat-Attenuation Index, or FAI.

Maya: Right. While basic PCAT analysis just looks at unadjusted mean attenuation, the FAI quantifies three-dimensional attenuation gradients, considering the non-linear influence of CT values. And to standardize this even further across different scanner vendors, they developed the FAI Score.

Sam: What makes the FAI Score different from just the raw FAI?

Maya: The FAI Score corrects the FAI values for those technical, biological, and anatomical factors we mentioned. It is then interpreted using age and sex specific nomograms. This harmonization is what creates a reproducible, platform-independent biomarker with real prognostic value.

Sam: And the prognostic value is huge. The paper states that elevated FAI Scores independently predict Major Adverse Cardiovascular Events, or MACE. And that is even after they adjust for things like high-risk plaque features, coronary calcium score, and traditional clinical risk factors.

Maya: Exactly. It is telling us something about the patient's risk that we simply cannot see by just looking at how blocked the artery is, or how much calcium is deposited there.

Sam: Now, the paper briefly touches on other imaging modalities before zeroing in on CT. Why not use MRI or PET scans to look at this inflammation?

Maya: They have their uses in research, but major limitations for everyday practice. PET-CT is great for molecular imaging of inflammation using tracers like FDG, but it has low spatial resolution, making it unreliable for tiny PCAT structures. Plus, it involves high radiation and high costs.

Sam: And MRI?

Maya: MRI provides radiation-free quantification of fat volume and composition, but again, it suffers from limited spatial resolution for detailed PCAT analysis, longer acquisition times, and high cost. CCTA remains the modality of choice because of its high spatial and temporal resolution, wide availability, and 3D volumetric data acquisition.

Sam: Okay, so we are sticking with CT. But actually tracing the borders of this tiny layer of fat around complex, branching coronary arteries manually? That sounds incredibly tedious and operator-dependent.

Maya: It is, which is why artificial intelligence is the critical bridge from research to clinical reality here. Traditional thresholding and semi-automated methods suffer from limited reproducibility. So the field has turned to deep learning, specifically convolutional neural networks, or CNNs.

Sam: The paper specifically mentions U-Net and its three-dimensional variants as the main framework for this medical image segmentation.

Maya: Yes, because they have a unique capacity to capture multiscale spatial features while maintaining very fine anatomical details. Advanced systems, like a hierarchical convolutional long short-term memory network and the 3D U-Net model, have achieved near-expert accuracy.

Sam: How accurate is near-expert? The document actually puts a number on it.

Maya: They achieved interclass correlation coefficients above 0.9 for various plaque types. These architectures are now being adapted to accurately identify the vessel wall and the surrounding adipose tissue.

Sam: This is a massive signal for anyone in the medtech and health AI space. The paper actually lists several commercial software platforms already incorporating deep learning for vascular wall and plaque segmentation. They name Autoplaque, CaRi-Heart, CaRi-Plaque, Cleerly, and HeartFlow Plaque Analysis.

Maya: It shows rapid commercial adoption. CaRi-Heart, for instance, specifically integrates fully automated segmentation of PCAT using deep learning. This gives us a reproducible, scalable solution to process massive datasets.

Sam: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Maya: And now, back to the document.

Sam: And speaking of massive datasets, you need them to train and validate these AI models so they do not fall apart in the real world. The paper highlights a couple of major registries doing this work.

Maya: Yes, they highlight CRISP-CT, which was a landmark study analyzing patients from Europe and the US, proving FAI Score is an independent predictor of mortality. But the really staggering one is the ORFAN registry.

Sam: ORFAN is massive. The paper says it includes over 100,000 CCTA scans, and they are aiming to reach 250,000 scans linked with long-term clinical outcomes.

Maya: That volume and diversity are essential. You have to validate these algorithms across different sexes, ethnicities, comorbidities, and scanning protocols. It proves the models are generalizable and gives us the statistical power to do health-economic modelling.

Sam: Which is exactly what health systems need before they will pay for this. They need to see that reclassifying a patient's risk based on their FAI score is actually cost-effective. But Maya, tell me this: if we identify this inflammation, can we actually do anything about it? Does the FAI score change if a patient gets treated?

Maya: That is perhaps the most exciting part of this paper for clinicians. PCAT imaging is not just for static risk prediction; it can monitor therapeutic response. The FAI score is a dynamic marker.

Sam: The paper details a specific prospective follow-up study looking at statins, right?

Maya: Yes. They looked at 52 statin-naïve patients with non-obstructive coronary artery disease. These patients received long-term high-dose statin therapy and had serial CCTA scans at baseline, around 1 year, and at 3 or more years.

Sam: And what did the serial scans show over those years?

Maya: There was a significant reduction in the FAI Score in the left anterior descending artery at both follow-ups. They also saw a decrease in non-calcified plaque volume and an increase in calcified plaque volume, which indicates the plaques were stabilizing. Total cholesterol, LDL, and HDL all improved.

Sam: So the statins didn't just lower cholesterol; they actively reduced the coronary inflammation, and the AI was able to read that reduction directly from the fat on the CT scan.

Maya: Exactly. But it is not just statins. The paper looks at novel anti-inflammatory agents and biologics as well.

Sam: Like the orticumab trial. This was a pilot randomized, double-blind, placebo-controlled trial.

Maya: Yes, in patients with psoriasis. They found that 15 weeks of orticumab, which is a monoclonal antibody targeting oxidized LDL, significantly reduced coronary inflammation as measured by the FAI Score, particularly in those with high baseline inflammation.

Sam: And they saw similar results in a larger observational study, right? 134 patients with psoriasis.

Maya: Correct. In that study, one year of biologic therapy targeting systemic immune pathways like TNF-alpha, IL-12/23, or IL-17 significantly lowered the FAI compared to matched controls. And importantly, this reduction was independent of lipid levels or traditional risk factors.

Sam: That is a huge point for pharma companies developing these biologics. The FAI score could potentially serve as an imaging biomarker to prove their drug is actively calming down the vascular inflammation driving heart disease.

Maya: There is also fascinating data on incidental anti-inflammatory effects. The paper mentions women with active breast cancer who underwent targeted thoracic radiotherapy. Prior to treatment, they had elevated FAI scores, likely driving elevated predicted cardiac mortality.

Sam: But two years after the radiotherapy, both their baseline FAI Score and their AI-predicted inflammatory risk actually declined.

Maya: Yes, independent of the local cardiac radiation dose. It highlights just how exquisitely sensitive PCAT is to local anti-inflammatory effects.

Sam: Now, the paper does point out that for some highly proven anti-inflammatory drugs, like colchicine and canakinumab, the imaging correlates are still under investigation. We know from trials like COLCOT, LoDoCo, and CANTOS that they reduce cardiovascular events, but we are still waiting on the direct FAI evidence.

Maya: The same goes for cardiometabolic agents like SGLT2 inhibitors and GLP-1 receptor agonists. Preliminary data is suggestive that they attenuate PCAT inflammation, but dedicated imaging trials are needed.

Sam: Okay, so FAI tracks dynamic, active inflammation. But what about permanent damage? If an artery has been inflamed for a decade, doesn't that leave a permanent mark on the tissue?

Maya: It does, and AI can see that too. The paper introduces a complementary imaging biomarker called the Fat Radiomic Profile, or FRP. While FAI looks at the dynamic, reversible changes in composition, FRP extends beyond simple attenuation.

Sam: Radiomics. That is where AI extracts hundreds of hidden texture and shape features from the image that the human eye could never see.

Maya: Exactly. FRP analyzes the radiomic texture in the area around the proximal right coronary artery. It detects persistent structural remodelling, including fibrosis and microvascular changes. These indicate more advanced, irreversible alterations.

Sam: So if FAI is a thermometer measuring the current fever in the vessel, FRP is looking at the scar tissue left behind from years of fevers.

Maya: A very helpful way to frame it. FRP has been shown to significantly improve the prediction of MACE beyond traditional risk factors, providing a non-invasive tool for assessing long-term adverse remodelling.

Sam: The paper also mentions a fascinating leap into something called radiotranscriptomics. They discuss a signature called the C19-RS.

Maya: This is cutting edge. The study introduces a novel AI-assisted radiotranscriptomic biomarker. It links the CT-derived PCAT features directly with underlying cytokine activity. The changes captured by the C19-RS signature effectively serve as molecular fingerprints quantifying the extent of vascular inflammation.

Sam: And they applied this to COVID-19 patients. The paper notes the signature was significantly elevated in these patients and demonstrated substantial prognostic value in predicting in-hospital mortality.

Maya: It suggests that PCAT-based vascular inflammation could be a key target for treatments like dexamethasone in those severe viral scenarios. It proves we can translate routine CT imaging into molecular-level biomarkers.

Sam: Alright, this all sounds like science fiction becoming reality, but we have to talk about the roadblocks. What is stopping every hospital from running this AI on every heart CT tomorrow?

Maya: The paper is very transparent about the technical limitations. First, manual or semi-automated segmentation introduces operator dependence. If two different technicians trace the fat slightly differently, you get different answers.

Sam: Which AI helps solve, but then the AI runs into scanner variability. Tube voltage, reconstruction kernels, scanner types... they all alter local attenuation values.

Maya: Yes. Plus, the presence of calcified or fibrotic plaque can also alter local attenuation values, making interpretation difficult in atherosclerotic segments.

Sam: And then there is the algorithm itself. The paper explicitly calls out explainability as a key limitation. These deep learning models function as black boxes. They give you a risk score, but limited insight into how specific image features influenced that outcome.

Maya: That opaqueness is a major hurdle. It complicates clinical interpretation, model validation, and regulatory approval. When a machine tells a cardiologist to completely change a patient's management plan, the cardiologist needs to understand why.

Sam: There is also a note on the business side. Because the FAI adjustments are proprietary, the details are not always fully accessible. The paper notes this might raise concerns for financial interest.

Maya: It is a valid concern, though the authors note that because proprietary AI models are regulated as medical devices, locked and subjected to independent regulatory approval processes, confidence in their reliability could be enhanced.

Sam: They also highlight the lack of large, well-annotated datasets for PCAT segmentation, noting most current studies depend on relatively small, homogeneous cohorts. This limits the ability of models to generalize across different populations.

Maya: Which is exactly why federated learning methodologies are the anticipated future direction. They will allow institutions to collaboratively train models without directly sharing patient data, protecting privacy while maximizing sample sizes.

Sam: To wrap up, the vision here is multiparametric prognostic modelling. AI won't just look at the fat. It will combine the deep learning plaque quantification, the FAI inflammatory measurements, machine learning survival models, and traditional clinical variables like Framingham or CAD-RADS scores into one comprehensive risk profile.

Maya: Exactly. AI is advancing cardiovascular imaging from single-modality analysis to comprehensive, systems-level disease characterisation. We are moving from a static anatomical map of a blocked pipe to a dynamic, biological readout of inflammatory risk.

Sam: One memorable takeaway for me: the fat around your heart isn't just sitting there. It is actively talking to your arteries. And when your arteries get sick, the fat changes its physical makeup in a way we can now measure and track over time.

Maya: It bridges mechanistic biology directly with bedside decision-making. But as the authors note, ongoing prospective studies and guideline incorporation are critical before this redefines everyday cardiovascular risk assessment.

Sam: That is a wrap on this deep dive into how AI and pericoronary fat analysis are transforming routine CT scans into dynamic biosensors. You can find a link to the source document in the show notes. And as always, a reminder to our listeners that this podcast is for informational purposes only and is not medical advice.