AI in Cardiac Imaging Has Arrived

Artificial intelligence holds the potential to streamline workflows and improve the performance of multiple cardiac imaging modalities, benefitting not only the physicians who specialize in these fields but also the patients they are treating.

“AI applications can span the spectrum from more basic machine learning–type models, which are implemented in risk prediction, for example, to deep learning models that we have used to analyze pixel-based data,” said Kate Hanneman, MD, MPH, FRCPC, associate professor and vice chair of research in the Department of Medical Imaging at the University of Toronto. “It can include interpretation of pictures to make a diagnosis for a specific disease classification, all the way to generative AI for generating a patient summary of a complex cardiac imaging report.”

Used effectively, integration of AI into the cardiac imaging workflow has the potential to improve efficiency, reproducibility, and standardization.

“AI can help improve diagnostic test performance, can decrease the number of misses, and minimize false positives,” Dr. Hanneman said. “There is tremendous potential.”

Application Types

To outline the potential applications of AI in cardiac imaging, Saurabh Jha, MBBS, MRCS, MS, associate professor of radiology at the Hospital of the University of Pennsylvania, grouped its use into several broad categories related to quantification, inference, and prognostication.

“Quantification is the AI trying to measure something—length, height, etc,” Dr. Jha said. AI software is already commonly used to measure ejection fraction. “Over the years, there were methods developed to do this semi-quantitatively, but now AI can essentially do the whole thing by itself.”

Available tools include LVivoEF, Real-Time EF, Low EF AI, and others. For example, Low EF was granted FDA clearance in 2024; the tool can detect low ejection fraction in 15 seconds using its “stethoscope.” Studies of the technology reported an area under the receiver operating curve of 0.85 for detection of left ventricular ejection fraction (LVEF) below 40%, with 84.8% sensitivity and 69.5% specificity.¹

In cardiac MRI, AI can take many of the time-consuming manual measurements, which are prone to inter-reader variability, and automate them. Dr. Hanneman characterized this use of AI for quantification as an “upstream” use.

What may be more exciting to physicians is the potential downstream uses of this technology, or what Dr. Jha categorized as “interpretation” or “prognostication.”

“This second-run use of AI would be to look at data and patterns and use that to predict disease,” Dr. Jha said. “For example, when we look at cardiac MRI, which is a very sophisticated form of imaging, there are patterns of abnormality. That isn’t measuring; now we are dealing with inference.”

Put another way, Dr. Jha said, you can’t measure a cat or a dog; you infer an animal is a cat or dog based on having seen so many cats or dogs.

Looking at inference, another recent study evaluated the accuracy of an AI system called PanEcho in automating echocardiogram interpretation with multitask deep learning. The system was developed using measurements from transthoracic echocardiography at Yale New Haven Health System hospitals and clinics and internally validated in a distinct cohort. PanEcho was highly accurate across geography and time from complete and limited studies. As an example, the system accurately estimated LVEF (mean absolute error: 4.2% internal; 4.5% external), as well as right ventricular systolic dysfunction, severe aortic stenosis, and other parameters.²

Another example is a recent study published in JACC: Cardiovascular Interventions that evaluated the ability of ChatGPT-4 to match human cardiologists in critical, complex medical decision-making. In 40 of 40 cases, the AI reached decisions that were identical to those of senior cardiology experts and aligned with European Society of Cardiology guidelines.³

A third area of potential in cardiac imaging is the use of AI for prognostication, Dr. Jha said.

“This would help give a better idea of what a patient’s clinical course is going to be,” Dr. Jha said.

For example, a study used machine learning techniques and personalized computational modeling to predict which patients with atrial fibrillation (AF) were most likely to experience AF recurrence after undergoing pulmonary vein isolation (PVI). The model, which used pre-PVI late gadolinium-enhanced MRI, predicted AF recurrence with a sensitivity of 82% and a specificity of 89%.⁴

New Exploration

This automation and efficiency at the point of care allow clinicians to save time for “high-caliber” activities, according to Partho P. Sengupta, MD, DM, FACC, FASE, chief of the Division of Cardiovascular Diseases & Hypertension at Rutgers Robert Wood Johnson Medical School.

“You can breathe a little bit. You can bring in a bit more of the humanistic quality interactions,” Dr. Sengupta said. “Burnout is a very true thing. These tools bring back the joy of work and make it more human.”

Dr. Sengupta said that the time AI has saved him has allowed him to spend more time in exploration and discovery. He has been asking whether “automation” can be “intelligence” and has been researching unsupervised learning.

“You look into patterns discovered by AI tools and are able to see and understand unique patterns that you could not have seen with your own eyes,” Dr. Sengupta said. “Instead of homogenizing everyone into one bucket, we are able to understand that some patients cluster into groups in different ways.”

Put more simply, Dr. Sengupta is using the lens of AI to explore the taxonomy of disease. “AI helps me spend less time on understanding the problem and more time on a solution,” he said. “Why is it that with cancer, [it] is always looked for in the early stage of disease, but in cardiology, we are often looking at the end? Why not find better ways to detect it earlier?

“Those paradigms have not been there yet,” Dr. Sengupta said, “but with AI we can start to do that.”

Unanswered Questions

The use of AI and available AI software is growing exponentially. Recently, a group of societies released a scientific statement giving an overview of the current landscape of AI applications in cardiac imaging. The statement discusses key steps in integrating the technology into a workflow but also focuses on potential drawbacks, including considerations related to environmental sustainability.⁵

“It is important that we link both its potential promise and its potential downsides,” Dr. Hanneman said.

Indeed, any excitement around the potential of AI must be tempered with a dose of reality related to unanswered questions that still surround it use.

“There are still major issues related to medicolegal implications, guidelines, and regulations,” Dr. Hanneman said. “These are not going to be easy to tackle and are very jurisdiction dependent and subject to political changes.”

Dr. Jha said that no discussion of AI is complete without talking about a host of ethical issues. Although not specific to AI in cardiac imaging, the “black box effect”—in which the internal workings of the AI system are unknown to the user—remains a key ethical question.

This element of “the unknown” reduces trust in AI and can affect adoption of AI programs or algorithms.⁶

Dr. Sengupta raised the unanswered question of “who is to blame” if mistakes involving AI occur. Currently, as pointed out by Priyom Bose, PhD, there is not an established line of responsibility “between healthcare providers, AI system developers, and regulators overseeing them regarding faulty judgments that harm patients.”⁷

“We are in an incredible inflection point,” Dr. Sengupta said. “These questions are all lingering and need to be resolved.”

Two other key questions, according to Dr. Jha, are “What is the usefulness of AI processing?” and “How do we know that it is useful?”

“Meaning, you can have a piece of information, and it is always good to know more stuff,” Dr. Jha said, “but that information only really makes a difference if you use it in a positive manner.”

Using AI for quantification seems to be capturing the more low-lying fruit, he said. The end goal will be using AI more for inference—something more useful—and that will require a lot more data processing.

Regardless of many of the remaining unanswered questions, Dr. Hanneman said that it is hard to see a future in cardiac imaging that does not involve AI.

“AI absolutely has the potential to help us,” she said. “As long as we use it judicially and intelligently.”