Evaluating Diagnostic Accuracy of an On-Site Deep Learning-Based CT-FFR Algorithm Using Invasive FFR as the Reference Standard
This independent single-center study validates that a deep learning-based on-site CT-FFR algorithm demonstrates good diagnostic accuracy, strong correlation with invasive FFR, and high reproducibility with processing times suitable for same-session reporting, though multicenter outcome-based validation is still needed.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
For decades, doctors have relied on two main ways to check if a heart artery is blocked. The first is a CT scan, which takes a detailed picture of the heart's plumbing. It is excellent at showing where the pipes are narrow, but it cannot always tell if that narrowness is actually stopping blood flow enough to cause trouble. The second method is an invasive procedure where a thin wire is threaded through the artery to measure pressure directly. This tells the truth about blood flow, but it requires a hospital visit, a needle, and a recovery period. For years, the goal has been to bridge this gap: to get the precise pressure data from the CT scan itself, without the need for the invasive wire. This would allow doctors to make life-changing decisions about treatment while the patient is still in the radiology suite.
A team of researchers at AZ Sint-Jan Brugge in Belgium recently put a new tool to the test to see if this vision is becoming reality. They evaluated a computer program that uses deep learning, a type of artificial intelligence, to calculate blood flow pressure directly from a standard CT scan. The software runs right on the hospital computer where the scan is viewed, promising results in minutes rather than days. The researchers wanted to know if this on-site program could match the accuracy of the invasive wire test, which remains the gold standard for diagnosis. They also wanted to see if different doctors would get the same results when using the tool, and how long it actually took to run the numbers.
To find the answers, the team looked back at records from 96 patients who had undergone both a CT scan and an invasive pressure test within three months of each other. The study involved two doctors with different levels of experience: a radiology resident and a senior board-certified radiologist. Both doctors analyzed the CT scans using the new software, which automatically maps the arteries and simulates how blood flows through them under stress. The software then predicted a pressure value for each blockage. The doctors were told where the blockages were located based on the invasive test, but they did not know the actual pressure results from that test, ensuring their judgment remained unbiased.
The results showed that the computer program was remarkably close to the invasive measurements. When the researchers compared the pressure values predicted by the software against the real measurements taken by the wire, the numbers lined up very well. The agreement was strong enough that the software correctly identified a significant blockage with an AUC of 0.80 in cases where the wire confirmed one was present. It also correctly identified when an artery was healthy with an AUC of 0.87 in those cases. This level of accuracy suggests the tool can reliably distinguish between arteries that need intervention and those that do not, a critical distinction for patient care.
Beyond raw accuracy, the study examined how consistent the tool was when used by different people. The two doctors, despite their difference in experience, produced very similar results when analyzing the same scans. When the resident doctor repeated the analysis on the same lesions weeks later, the results were nearly identical to the first time. This consistency is vital for a medical tool, as it means the outcome depends on the technology and the patient's anatomy, not on the mood or skill level of the specific doctor using it. The software also proved fast. In a subset of patients where the time was tracked, the analysis took an average of 10 minutes and 33 seconds. This speed is fast enough to fit into a standard workday, allowing a doctor to review the scan, run the pressure calculation, and discuss the results with the patient all in the same session.
The researchers noted that the tool is not perfect. Most of the errors occurred in cases where the pressure was right on the borderline between healthy and diseased, a zone where even the invasive wire test can be difficult to interpret. In these tricky cases, the software sometimes guessed slightly too high or too low. The study also highlighted that the software is not fully automatic; it requires a doctor to check the computer's map of the arteries and make small edits if the image is unclear, such as when heavy calcium deposits obscure the view. This manual step means that the quality of the final answer still depends partly on the human operator.
Ultimately, this study provides strong evidence that deep learning can bring invasive-level precision to a non-invasive scan. The tool demonstrated good diagnostic accuracy, strong agreement with the gold standard wire test, and the speed necessary for real-world use. While the researchers caution that larger studies across different hospitals are needed to confirm these findings for the general population, the results suggest a future where the decision to open a blocked artery can be made with greater confidence and less delay, all while the patient remains in the comfort of the imaging center.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.