Development of an automated, reliable, and clinically meaningful artificial intelligence (AI) tool for diagnosing cardiac disease from conventional cardiovascular magnetic resonance (CMR) images

Heart disease remains the single greatest threat to human life worldwide, claiming millions of lives every year. To understand what is happening inside a beating heart, doctors rely on a powerful imaging technique called cardiovascular magnetic resonance, or CMR. Unlike a standard X-ray that shows a flat shadow, this technology creates detailed, moving pictures of the heart's muscle, its pumping action, and the health of its tissue. It can reveal whether the heart muscle is thickened, stretched, or damaged by a lack of blood flow. However, reading these complex images requires a high level of specialized training and years of experience. A single scan contains hundreds of moving frames and different types of tissue data, making it a heavy cognitive load for even the most skilled physician. In many parts of the world, there simply are not enough experts available to interpret every scan accurately and quickly. This gap between the need for diagnosis and the availability of experts has driven researchers to ask if a computer could learn to read these images with the same care and precision as a human specialist.

A team of researchers at the University Hospital Münster in Germany has taken a significant step toward answering that question. They developed a new artificial intelligence system designed to look at CMR images and automatically determine if a patient has a specific type of heart disease. Rather than building a computer program from scratch, they started with "foundation models." These are massive, pre-trained computer systems that have already learned to recognize patterns in millions of images and texts. Think of these models as a student who has already read every book in a library and seen every picture in a gallery; the researchers' job was simply to teach this student how to apply that vast knowledge specifically to heart scans. The team focused on five distinct categories: a healthy heart, and four common forms of heart muscle disease known as hypertrophic cardiomyopathy, dilated cardiomyopathy, ischemic cardiomyopathy, and cardiac amyloidosis.

The first major hurdle in creating such a tool is not teaching the computer to see, but teaching it what to look for. In a typical hospital, patient records are written as long, narrative paragraphs by doctors, describing what they see in the images. To train an AI, these stories must be converted into clear labels, a process that usually requires a human to read thousands of reports and type out the diagnosis. This is slow, expensive, and prone to human error. The researchers solved this by using a different kind of artificial intelligence: large language models. These are the same types of systems that can write essays or answer questions. The team set up a secure, local system where these language models read the German medical reports, translated them into English, and then extracted the specific diagnosis for each patient. To ensure accuracy, they used three different language models to read every report. If at least two of the three agreed on the diagnosis, that label was accepted. This automated process allowed them to curate a massive dataset of nearly 1,000 patient cases without needing a human to manually label every single file.

Once the data was prepared, the researchers turned to the image analysis. They fed the AI three different types of heart images: a short-axis view that looks like a stack of slices through the heart, a four-chamber view that shows the heart from the side, and a special scan called late-gadolinium enhancement that highlights scarred or damaged tissue. They fine-tuned three different vision foundation models on this data. The first stage of training taught the models to recognize the general features of the heart, and the second stage refined their ability to distinguish between the five specific disease categories. The researchers then tested the system on a completely separate group of over 1,000 patients whose diagnoses were confirmed by experienced cardiologists. This independent test ensured that the AI was not just memorizing the training data but was actually learning to recognize the diseases.

The results were striking. When the AI looked at the images alone, it achieved a high level of accuracy in identifying the different heart conditions. For the disease known as cardiac amyloidosis, the system correctly identified it in nearly every case, and for the thickened heart muscle condition, it performed with similar success. However, the researchers found that the system worked even better when it combined the insights from all three image types and all three different AI models. By letting the models vote on the final diagnosis, they created a "consensus" that was more reliable than any single model could be on its own. In this combined approach, the system reached its highest accuracy for the thickened heart condition and cardiac amyloidosis, correctly distinguishing them from a healthy heart with a high degree of confidence. It also performed well on the other two disease types, though it found it slightly more difficult to tell the difference between a heart stretched by disease and a heart damaged by a lack of blood flow, as these conditions can sometimes look very similar on a scan.

Beyond just getting the right answer, the researchers wanted to know if the AI was fair and if it was looking at the right parts of the heart. They checked the system's performance across different ages and between men and women. The results showed that the AI worked equally well for both sexes, though there was slightly more variation in how it performed across different age groups. This is likely because certain heart diseases are much more common in older adults, making the data for younger patients with those specific conditions harder to learn from. To understand what the AI was actually seeing, the researchers used a visualization tool that highlighted the specific areas of the image that influenced the decision. The computer consistently focused its attention on the heart muscle itself, particularly the left ventricle, and ignored the surrounding structures. This confirmed that the system was making its decisions based on the actual anatomy of the heart, rather than guessing based on unrelated background details.

The study concludes that this approach offers a viable path forward for clinical practice. By combining automated report reading with advanced image analysis, the team created a tool that could potentially support doctors who do not have years of specialized training in reading heart scans. The system does not replace the doctor; instead, it acts as a second pair of eyes that can quickly sort through complex data to suggest a diagnosis. The researchers acknowledge that their work is a proof of concept based on data from a single hospital, and that future studies will need to test the system on a wider variety of patients and with more types of heart disease. They also noted that the current system relies on the images alone and does not yet incorporate other clinical data like blood tests or patient history. Nevertheless, the study demonstrates that with the right combination of automated data preparation and modern artificial intelligence, it is possible to build a system that is accurate, interpretable, and capable of supporting the correct interpretation of heart scans in a clinical setting.