Mode
Text Size
Log in / Sign up

Review of fine-tuned AI model for coronary artery disease image analysis and report generationAI Can Now Read Heart Scans—But Is It Ready to Help Your Doctor?

AI-generated summary of the cited source, checked by automated accuracy review. How we work

Key Takeaway
Consider AI model metrics for CAD analysis but note lack of safety and comparative data.

This publication is a development and evaluation study review focusing on a fine-tuned InternVL2-4B model with Low-Rank Adaptor weights. The analysis utilized four unique datasets comprising 1987 patients to evaluate the model's capabilities in coronary artery disease contexts. The primary outcomes assessed included stenosis detection, anatomy labelling, and report generation, with secondary outcomes not reported in the source material.

Key performance metrics were derived from the evaluation. For stenosis detection, the model achieved a precision of 0.56, a recall of 0.64, and an F1-score of 0.60. Regarding anatomy segmentation, the weighted precision was 0.50, recall was 0.43, and the F1-score was 0.46. The report generation task yielded an accuracy of 0.42, a negative predictive value of 0.58, and a specificity of 0.52. No absolute numbers, p-values, or confidence intervals were reported for these outcomes.

The review highlights potential practice relevance, such as streamlining angiogram interpretation to rapidly provide actionable information to guide management. It also suggests the model could support care in resource-limited settings and aid in auditing the appropriateness of coronary interventions. However, the study did not report adverse events, serious adverse events, discontinuations, or tolerability data. Furthermore, no comparator was reported, and the study phase was not reported.

Given the absence of comparative data and safety information, the clinical applicability remains uncertain. The authors do not provide specific limitations beyond the lack of reported comparators and safety data. Clinicians should interpret these performance metrics with caution until further validation is established in broader, randomized contexts.

A New Tool for a Common Problem

Imagine sitting in a hospital room, waiting for news about your heart scan. The doctor is busy, and the results take time. Now, imagine an AI system that looks at your heart scan and writes a detailed report in minutes.

This is what a new study is testing. Researchers built an AI tool that can read coronary angiograms—scans that show blockages in the heart’s arteries. The goal is to help doctors make faster decisions.

Coronary artery disease is a leading cause of death worldwide. Doctors often use angiograms to see how bad the blockages are. But reading these scans takes time and skill. In busy hospitals or remote areas, there may not be enough experts to read them quickly.

This delay can mean waiting days for results. That’s stressful for patients and can slow down treatment. An AI tool that reads scans and writes reports could help speed things up.

The Old Way vs. The New Way

Traditionally, doctors or trained technicians look at each scan frame by frame. They measure the blockages and write a report by hand. This process is accurate but slow.

But here’s the twist: AI can now look at thousands of scan images in seconds. The new tool doesn’t just spot blockages—it also labels the arteries and writes a full report. This could save time and reduce errors.

Think of the AI as a smart assistant for doctors. It uses a “vision-language model,” which means it can see images and understand language. It’s like a translator that turns pictures into words.

The AI was trained on 20,000 scan images from nearly 2,000 patients. It learned to spot blockages, label the arteries, and write a report. The system uses a technique called “low-rank adaptor weights,” which is like adding a small upgrade to a computer program instead of building a new one from scratch.

Researchers tested the AI on scan images from four different datasets. They checked how well it detected blockages, labeled arteries, and wrote reports. The study was published on medRxiv in April 2026.

The AI detected blockages with 60% accuracy. For labeling arteries, it was about 46% accurate. When writing reports, it was 42% accurate.

These numbers may seem low, but they are a start. The AI was better at spotting major blockages and major arteries. For example, it was more accurate with the left anterior descending artery—a key artery in the heart.

But here’s the catch: human doctors are still more accurate. The AI is not ready to replace them. Instead, it could act as a helper, flagging issues for doctors to review.

This doesn’t mean this treatment is available yet.

Researchers say this AI tool could be useful in places with few specialists. It could also help doctors double-check their work or audit whether a heart procedure was needed. But they stress that more testing is needed before it can be used in real hospitals.

This AI tool is still in the research phase. It’s not available in hospitals yet. If you’re waiting for a heart scan result, talk to your doctor about the process. They can explain how long it takes and what to expect.

The study has some weaknesses. The AI was tested on a limited set of images, and its accuracy is not yet at human levels. It may also struggle with unusual cases or low-quality scans. More research is needed to improve its performance.

Next, researchers plan to test the AI in real hospitals. They will also work on making it more accurate and reliable. If successful, this tool could be available in clinics within a few years. For now, it’s a promising step toward faster, more efficient heart care.

Study Details

Sample sizen = 1,987
EvidenceLevel 5
PublishedApr 2026
View Original Abstract ↓
Coronary artery disease is a leading cause of morbidity and mortality. Invasive coronary angiography is currently the gold standard in disease diagnosis. Several studies have attempted to use artificial intelligence (AI) to automate their interpretations with varying levels of success. However, most existing studies cannot generate detailed angiographic reports beyond simple classification or segmentation. This study aims to fine-tune and evaluate the performance of a Vision-Language Model (VLM) in coronary angiogram interpretation and report generation. Using twenty-thousand angiogram keyframes of 1987 patients collated across four unique datasets, we finetuned InternVL2-4B model with Low-Rank Adaptor weights that can perform stenosis detection, anatomy labelling, and report generation. The fine-tuned VLM achieved a precision of 0.56, recall of 0.64, and F1-score of 0.60 for stenosis detection. In anatomy segmentation, it attained a weighted precision of 0.50, recall of 0.43, and F1-score of 0.46, with higher scores in major vessel segments. Report generation integrating multiple angiographic projection views yielded an accuracy of 0.42, negative predictive value of 0.58 and specificity of 0.52. This study demonstrates the potential of using VLM to streamline angiogram interpretation to rapidly provide actionable information to guide management, support care in resource-limited settings, and audit the appropriateness of coronary interventions.
Free Newsletter

Clinical research that matters. Delivered to your inbox.

Join thousands of clinicians and researchers. No spam, unsubscribe anytime.