«Today, artificial intelligence is truly changing the way we diagnose diseases. In recent years, science has made enormous progress in automated image analysis, and these advances are extremely important for medical applications and for innovations in radiology — or more precisely, diagnostic imaging.»

This is how Antonio Esposito, Deputy Scientific Director and Head of the Radiology Unit for Personalized Medicine and the Cardiovascular Radiology Unit at IRCCS San Raffaele Hospital, and Full Professor at Vita-Salute San Raffaele University, explains it.

We interviewed him on the occasion of World Heart Day, since Professor Esposito has recently developed an AI model designed to optimize the diagnosis of coronary atherosclerotic disease, the underlying condition that causes heart attacks.

Coronary Disease and Myocardial Infarction

Coronary artery disease (which affects the vessels supplying blood to the heart) is caused by atherosclerosis—a pathological thickening of the arterial walls, resulting from various factors that lead to the formation of plaques that grow due to fat accumulation, inflammation, and other pathophysiological processes.

When these atherosclerotic plaques rupture or detach, they can suddenly block the coronary arteries, interrupting blood flow to the heart and causing the death of part of the cardiac muscle—this is what we call a myocardial infarction (heart attack).

“Today, a patient suspected of having coronary disease undergoes a CT scan (Computed Tomography) to observe the state of the vessels and measure the degree of stenosis (that is, how much they have narrowed). This assessment helps determine whether the patient has ischemia, a reduction in blood flow to the heart during periods of increased demand, such as during everyday physical effort.

This is also important for understanding the likelihood that the patient might have a heart attack in the near future. However, while measuring stenosis is very useful, it is not sufficient for accurate prediction. One of my research projects aimed to use AI to automatically extract additional quantitative information from CT scans — beyond stenosis — data that can reveal much more about the pathophysiological processes occurring within the artery walls. AI can extract and analyze this information to predict the risk of heart attack for each individual patient,” explains the Professor.

Training Artificial Intelligence to Read Coronary CT Scans

To achieve this, Esposito and his team trained an AI model to interpret coronary CT images in an innovative way. This virtual “eye” recognizes a set of features within the images that together form a signature of the individual’s risk.

Just as a handwriting expert examines a signature to determine its authenticity, the AI analyzes visual patterns to estimate the likelihood that a given person will suffer a heart attack in the coming years. The results have been exceptional, with the method demonstrating high accuracy in predicting individual risk.

The software has already been internally validated. Once broader validation is completed, it can be certified as a tool to support physicians in everyday clinical practice—allowing for personalized therapies based on each patient’s unique characteristics and risk profile. This would improve both patient outcomes and healthcare resource management.

Based on its prior training, the model estimates the risk of heart attack for each new patient, enabling more targeted and informed decisions.

From Detail to the Whole Picture: What the Model Sees

“Most currently available software analyzes coronary images by dividing them into small segments and extracting volumetric measurements. Our approach is entirely new. It uses the full CT dataset, analyzing subtle image variations that might seem insignificant but actually reflect processes occurring within the heart—processes involving moving, interacting, and dying cells and molecules. These details, invisible to the naked eye, are crucial because they tell us what may happen in one, two, three, or four years, allowing us to intervene and change the clinical course of each individual,” Esposito says.

To better understand this, imagine observing three paintings depicting the same hilly landscape, each created by a different artist. Each painting will have distinct brushstroke shapes, directions, and colors that reveal the artist’s unique hand. The AI, by observing the painting as a whole, can read these details and tell us who the painter is.

“What our model does,” Esposito continues, “is observe the whole picture—the CT image of the heart and coronary arteries—and detect the subtle nuances that represent the signature of biological processes occurring at the cellular and tissue level, processes that underlie a future heart attack. Based on what it has learned, the model can then say: given this pattern, there is a certain probability that this patient will have a heart attack.”

Clinical Impact and Future Prospects

The model developed by Esposito and his collaborators is ready and has so far performed very well. In the coming months, additional validation studies will be conducted. Its potential impact is significant.

Currently, clinicians can identify patients at very high or very low risk reasonably well, but most people fall into the intermediate-risk category. Within this large group, however, are individuals whose true risk is either very low or very high.

As a result, some patients are over-treated, receiving unnecessary therapy, while others are under-treated, missing out on essential preventive care—and they are the ones who often go on to have heart attacks.

“Our model has shown excellent performance even in this ‘in-between’ group, accurately distinguishing those at genuinely low risk from those at high or very high risk. This means giving therapies only when needed, to the right patient, avoiding unnecessary medications for some while preventing heart attacks in others,” the Professor explains.

It may sound like science fiction, but it’s simply mathematics, computer science, and statistics in the service of medicine.

Cultivating Critical Thinking and Awareness in the Age of AI

Artificial intelligence is a valuable aid, but it cannot replace the physician, who remains responsible for making the most appropriate diagnostic and therapeutic decisions.

Therefore, these new tools must be studied, understood, and used consciously. This is especially true for young doctors in training—students and residents—who must learn how to use AI but must not depend on it. They need to cultivate and maintain their critical thinking and decision-making abilities.

“My goal,” Esposito concludes, “is to help students develop that human critical sense and awareness that are essential to turning any radiological image into an accurate diagnosis.

Any doctor beginning to use an AI-based tool must carefully study its technical characteristics, possible applications, design assumptions, and training data. This ensures, as much as humanly possible, a conscious and critical use of these important tools.”