ACC Journal Club | AI in Cardiovascular Imaging: A Transatlantic Perspective on the Future of Cardiology

How far has AI already entered daily cardiovascular imaging workflows in your environment – and where do you see the most immediate clinical benefit: acquisition, interpretation, or decision support?

Meder: Since we have been confronted with AI, we have started to examine existing workflows more closely and to reassess the accuracy of diagnostic and therapeutic decisions. In this sense alone, AI has already helped to question, analyze, and improve processes. In diagnostic modalities, AI is often implemented in a “device-centered” manner, as integration into closed systems is easiest for vendors. As in many other institutions, however, implementing existing – and sometimes very good – AI solutions remains a cumbersome process. Ultimately, a great deal of efficiency is still being lost.

Cavalcante: AI is a routine presence in our daily CV imaging workflow. Whether it is measuring volumes and EF on echo and/or CMR, or identifying high-risk features for HFpEF or quantifying atherosclerotic plaque in coronary CTA, AI provides value by reducing the time spent on repetitive post-processing tasks. While these examples relate mainly to post-processing, we look forward to expanding the role of AI in better standardizing imaging acquisition, enabling rapid longitudinal interval comparisons, and supporting decision-making – not only in choosing the right diagnostic imaging test, but also in guiding the ordering provider based on test results.

How do regulation, data infrastructure, and reimbursement policies influence the speed of AI adoption in your healthcare system?

Meder: The landscape of reimbursement is bleak, which initially limits the financing of AI solutions. The focus therefore lies on quality and efficiency gains. While these can be demonstrated in a business plan – and many studies show that a medical AI solution often pays for itself within the first year – this requires a comprehensive redesign of workflows.

Cavalcante: At Allina Health Minneapolis Heart Institute, we have formed a multidisciplinary Innovation and Technology committee with several stakeholders including VP of Operations, VP of Strategy, Imaging Section Head, Healthcare Delivery Innovation Section Head, and additional key administrative leaders to specifically address the evaluation and integration of AI into our practice. The main goal of this committee is to regularly assess our workflows within cardiac imaging and to identify where potential AI solutions could provide improved throughput, patient and provider experience, and automation. After internal alignment and prioritization based on the urgency of each clinical section, presentation of solutions and further discussions can lead to important strategic partnerships that are mutually beneficial.

Which AI applications have already entered routine clinical practice in your environment, and where do you expect the next steps of implementation?

Meder: MRI, CT, echocardiography, and ECG are common examples. Large language models are currently used primarily in research, with a clear focus on reducing physician workload.

Cavalcante: One AI application we use on a daily basis is coronary CTA – either for patients in whom plaque quantification by Cleerly Labs helps inform treatment decisions (e.g., intensification of preventive measures or medications), or by using CT-FFR to determine the need for invasive coronary angiography and to plan PCI, including estimation of expected benefit using the HeartFlow PCI Planner. This allows us to personalize patient care and to guide practitioners toward appropriate decisions based on diagnostic imaging. We are also exploring CT-guided PCI together with key interventional cardiology colleagues (Drs. Yader Sandoval and Manos Brilakis), evaluating whether this planning can be used not only to triage case complexity to the appropriate treatment site and operator, but also to improve outcomes beyond the current standard of care.

Can we trust automated measurements – and how should clinicians maintain oversight?

Meder: We already work with many forms of human intelligence in hospitals. From this, we have learned that plausibility checks, sampling, and cross-verification are essential pillars of safe care. The same applies to AI-based medical devices, which require comparable standards and ongoing training.

Cavalcante: As I teach my fellows and junior colleagues: always verify before you trust. This is particularly important in cardiac imaging, as the training datasets used to derive AI algorithms may differ from the individual patient being analyzed. Technological advances to improve imaging quality should be pursued alongside comprehensive staff training to maximize equipment capabilities and match them to patient complexity.

Will AI in imaging be accessible to everyone, or will expensive infrastructure remain a barrier?

Meder: To everyone – without exception. Just don’t ask me for a timeline.

Cavalcante: Increased competition across multiple vendors within specific areas of cardiac imaging is beneficial – not only by continuously improving product value, but also from a reimbursement perspective. Over time, I expect continued growth of AI in cardiac imaging, with broader adoption not only in academic centers but also in large integrated group practices. The biggest barrier we currently face is the speed at which IT departments can complete thorough evaluations and resource assessments after cybersecurity clearance – a process that can take several months.

How should training evolve to ensure physicians remain central in interpretation and ethical use of AI tools?

Meder: For me, the key is “humane” medicine. Technology is valuable as long as the focus remains on the patient. The German Cardiac Society’s eCardiology group has recently published a position paper on digital education, emphasizing the need to teach digital competencies. Current meetings offer numerous initiatives to train the next generation responsibly, courageously, and with a human-centered mindset.

Cavalcante: Cardiology training already incorporates AI, and earlier exposure is beneficial. However, AI should be framed as an adjuvant that augments provider knowledge. It can help address specific questions and uncommon or challenging scenarios where management may follow multiple paths. I frequently use the OpenEvidence app/website, which I believe has transformed the medical field. Free, rapid consultation with high-quality summaries of relevant references is now possible. To me, this enables continuous knowledge improvement and evidence-based patient care. Again, AI augments – but does not replace – clinical reasoning, shared decision-making, and bedside skills.

High-quality, annotated imaging data are the foundation of robust AI models. What strategies or networks exist to enable secure, large-scale data sharing?

Meder: Many researchers are engaged through the German Cardiac Society, the German Centre for Cardiovascular Research, and registries. Prof. Engelhardt strongly promotes federated learning – allowing algorithms to travel rather than data. Often, more data already exist across sources than we realize.

Cavalcante: Indeed, high-quality, well-annotated cardiac imaging data are the foundation of robust and clinically trustworthy AI models. However, single-institution datasets lack the scale, diversity, and generalizability required for real-world deployment. Minnesota is uniquely positioned to lead secure, large-scale cardiac imaging data collaboration by leveraging federated networks, strong governance frameworks, and advanced imaging infrastructure. Without secure, multi-institutional collaboration, AI models risk bias, poor generalizability, and limited clinical impact. Successful cardiac AI depends as much on trust, governance, and collaboration as it does on technology.

Many AI algorithms show strong technical performance but limited prospective validation. How do you address validation, transparency, and clinician trust?

Meder: The trend is clear: from initial single-center observations toward multicenter, randomized studies assessing whether AI is safe and superior to best medical practice. Han et al. (The Lancet Digital Health 2024) showed that in 85 of 86 randomized studies, AI was not inferior, and in 65 studies provided significant benefit.

Cavalcante: This is a critical part of any AI tool entering the market. A minimum requirement is obtaining the necessary FDA clearance. Equally important is understanding the patient population used for algorithm development, the rigor of validation, and the current limitations of the released product. Ongoing feedback between users, institutions, and vendors is essential to ensure quality standards are met.

Over the next 5–10 years, which combination of technologies will most strongly impact patient care – and what role will human expertise continue to play?

Meder: I see great potential for broad improvement. AI-enabled ECGs in primary care may correctly detect STEMI, predict hypokalemia, or trigger evaluation for structural heart disease – leading to faster, more accurate diagnostics across sectors. Language models may also help patients better understand complex medicine.

Cavalcante: The combination of multimodal imaging and predictive analysis is particularly exciting to me, especially in the field of structural transcatheter interventions. Timely diagnosis combined with thorough phenotypic assessment may allow us to truly apply the principle of “the right patient, at the right time, for the right therapy.” Our role as physicians will remain critical in identifying educational and practice gaps and in continuously evaluating where and how AI implementation can meaningfully address these challenges.

To the overview page international content