AI in Cardiac Imaging: Structural vs. Functional Evidence

Cardiovascular AI is moving fast, with a ton of new methods for figuring out patient risk and managing disease. But the real nut to crack is generating proof that any of it works. This is especially true for quantifying coronary plaque, where the promise of AI is getting a tough shakedown against old-school clinical standards. The whole field is split between two camps, structural analysis and functional assessment, and that divide dictates how these AI tools get built, tested, and eventually used on patients.

The CERTAIN Trial and the Centrality of Plaque Burden

Any discussion about AI in coronary CT angiography (CCTA) usually starts with the structure of coronary plaque. Cleerly has built its whole CCTA program around this idea, anchoring it to the multicenter CERTAIN trial publication on clinicaltrials.gov. That study, published in January 2024, was a big deal. It showed that when you use AI to quantify plaque, you change how doctors manage patients. The model moves from just chasing stenosis to getting a full picture of atherosclerotic burden. Cleerly’s pushing this forward with things like the CONFIRM2 Registry, a huge international study looking at the prognostic value of its AI-based plaque work. The shift to focusing on plaque burden over simple luminal narrowing is a complete reorientation of how we think about cardiac risk. Why? Because we now know that non-obstructive plaque can be incredibly vulnerable and cause a heart attack. The CERTAIN trial showed that AI can get us past fuzzy qualitative reads and give us hard, repeatable numbers on plaque volume, composition, and other features. This kind of detail helps a clinician get way more precise with risk stratification, which then informs everything from lifestyle and medication advice to how often the patient needs to come back for a check-up. Trials like CERTAIN set a very high bar for clinical utility. Any new AI has to prove its technical accuracy, sure, but it also has to show it can actually change patient outcomes for the better.

Structural Analysis Versus Functional Assessment in CCTA

The CCTA world is basically split down the middle in how it approaches coronary artery disease: you’re either analyzing the structure of the plaque or you’re assessing the function of the blood flow. Cleerly’s program is all about the structural analysis, while HeartFlow corners the functional market with its CT-derived fractional flow reserve (CT-FFR) tech. Cleerly’s structural approach involves pulling a huge amount of detail about atherosclerotic plaque out of CCTA images. The AI measures total plaque volume, low-attenuation plaque (the bad stuff), calcified plaque, and fibrous plaque, which all feed into a patient’s total atherosclerotic burden and risk profile. The whole idea is that the plaque’s composition and shape are powerful predictors of future heart attacks, regardless of how much it’s actually narrowing the artery. AI algorithms are just better at this than a human eye, segmenting and characterizing plaque with a detail and consistency that people can’t match Review article on AI for plaque characterization in CCTA. HeartFlow, on the other hand, is all about function. Their software applies computational fluid dynamics to the CCTA scan to calculate FFR, which is a direct measure of how much a blockage is restricting blood flow. FFR is the established physiological marker for ischemia, and CT-FFR is a non-invasive way to find the specific lesions that are hemodynamically significant enough to need a stent. Most of the evidence for CT-FFR has shown that it helps avoid unnecessary invasive angiograms and makes diagnosis more accurate for patients with stable chest pain. These two methods, structural and functional, give a clinician vital information. The structural analysis gives you the big picture of the disease burden and long-term risk, while the functional assessment tells you the immediate hemodynamic impact of a specific lesion. They answer different questions. They are complementary, offering different views for evaluating coronary artery disease. Putting both structural and functional data together, using AI to do the heavy lifting, is where complete CCTA interpretation is headed.

Parallel Evidence Designs in Adjacent Modalities

The kind of tough evidence standards we’re seeing in CCTA aren’t isolated. You see similar patterns popping up in other parts of cardiac imaging and even completely different medical specialties, all pointing toward a consensus on what it takes to get an AI into the clinic. Take Ultromics, which went a different route by focusing on echocardiography. Their EchoGo platform has a string of FDA clearances, for example, EchoGo Heart Failure helps spot heart failure with preserved ejection fraction (HFpEF), and EchoGo Amyloidosis is for detecting cardiac amyloidosis. The platform’s very first FDA clearance, FDA 510(k) database entry for EchoGo, was based on proving the AI could accurately measure left ventricular ejection fraction (LVEF) and global longitudinal strain (GLS), key metrics of cardiac function. Subsequent clearances just built on that foundation. The evidence path for EchoGo looked a lot like Cleerly’s: rigorous validation against expert human reads and, critically, against actual clinical outcomes. It just shows that no matter the imaging type, the demand for strong clinical evidence is the same. Using AI to automate and standardize complex echo measurements promises to cut down on variability and speed up diagnosis, just like AI in CCTA is doing for plaque. If you want a real benchmark for evidence standards, look outside of cardiology at Paige AI. In September 2021, they got the first-ever FDA approval for an AI to perform primary diagnosis in digital pathology, specifically for detecting prostate cancer on whole slide images FDA press release on Paige AI approval. That approval completely reset the bar. It showed that an AI could get cleared as a primary diagnostic tool, not just a decision support add-on. The validation Paige needed was massive, involving huge datasets and proving it was as good as (or better than) human pathologists. For the emerging field of cardiac pathology, the message is obvious: any AI is going to need to show similar diagnostic accuracy and clinical benefit to get adopted and approved.

The Shared Evaluation Base for Coronary CT

As cardiac AI develops, you can see a common set of evaluation criteria taking shape, especially around coronary CT. For many of these new AI programs, coronary CT angiography and its risk prediction models have become the common ground for evaluation, whether the tool focuses on plaque structure or blood flow. This shared evaluation base means any AI has to show more than just fancy image analysis. It has to clearly affect how patients are risk-stratified and how doctors make decisions. The one thing that connects all these different AI applications is their ability to accurately identify patients at high risk of a major adverse cardiovascular event (MACE). It doesn’t matter if the AI is quantifying plaque, assessing perfusion, or predicting events five years out. Its value is measured by whether it leads to better patient outcomes and more efficient care. The high standards from trials like CERTAIN, the intense validation needed for FDA clearances like those for EchoGo and Paige AI, and the constant watch from regulatory bodies all create a strong framework for evidence. This framework makes sure that AI in cardiology is a clinically validated tool that sharpens diagnosis and guides treatment to improve cardiovascular health. The future of cardiac AI will be determined by which companies can best navigate this complex evidence field, proving both their technical chops and their undeniable clinical worth.