Chest pain is one of the most frequent reasons for visits to emergency and outpatient departments, yet diagnosing underlying coronary artery disease remains complex and resource-intensive. Coronary CT angiography (CCTA) has emerged as a preferred diagnostic method for patients with intermediate risk of coronary artery disease, especially in the absence of prior cardiovascular events. However, reliance on expert interpretation has limited widespread implementation. Recent advances in artificial intelligence offer an opportunity to streamline CCTA workflows. A recent study published in Radiology assesses the diagnostic accuracy of an on-premise AI tool for automated coronary artery calcium scoring (CACS) and CCTA interpretation, using the CAD-RADS 2.0 classification system, comparing it against expert human analysis.
AI Performance in Coronary Artery Disease Assessment
The retrospective study analysed 1,041 CCTA scans from 1,032 patients obtained across four different scanners from three vendors. The on-premise AI software generated results in less than five minutes, producing CACS, CAD-RADS categories and plaque burden scores. AI agreement with expert readers was substantial to near-perfect across multiple parameters. For CAD-RADS classification, the agreement reached a weighted κ of 0.73, and the AI demonstrated a high area under the ROC curve (AUC = 0.90) for detecting moderate (≥3) and high-grade (≥4A) stenosis. The tool showed a particularly strong negative predictive value (NPV) of 98% for CAD-RADS ≥4A, indicating its strength in ruling out obstructive disease. However, the positive predictive value (PPV) remained modest at 39%, highlighting a limitation in detecting high-risk patients without supplementary expert review.
Must Read: Ultra-low-dose Coronary CT with Deep Learning Reconstruction
In plaque burden scoring, AI reached almost perfect agreement with expert readers for CACS-derived scores (κ = 0.97) and substantial agreement for segment involvement score (SIS)-derived assessments (κ = 0.79). While AI occasionally over- or underestimated the CAD-RADS category, a one-category tolerance range brought its accuracy to 87%. Notably, motion artefacts and scanner-specific variations emerged as primary contributors to misclassification.
Benefits of On-Premise Deployment and Workflow Integration
Unlike cloud-based AI solutions, this on-premise implementation allows immediate processing, preserving patient data security and reducing turnaround time. Its seamless integration into clinical environments enhances triage, prioritisation and decision-making, especially in settings with limited radiology expertise or high imaging volumes. The system’s performance was consistent across different CT scanner models and acquisition modes, demonstrating its adaptability to diverse clinical infrastructures.
Additionally, the AI module’s ability to automate CACS and plaque burden assessment contributes to standardisation and efficiency, reducing interobserver variability and manual workload. In emergency departments, such automation could accelerate discharge decisions for low-risk patients and ensure prompt attention for those requiring intervention. Though the system’s lower PPV necessitates human oversight, its high NPV makes it a powerful adjunct tool for preliminary exclusion of disease.
Study Strengths, Limitations and Clinical Implications
This study stands out by evaluating all components of the AI tool—CACS, CAD-RADS classification and plaque burden scoring—on a large and diverse patient sample. It used a multivendor and multi-scanner setup reflective of real-world conditions, addressing the limitations of earlier validation studies. However, certain constraints should be noted. Expert interpretation, not invasive angiography, served as the reference standard. Also, adjudication was limited to cases with discrepancies of more than one CAD-RADS category, possibly leaving minor misclassifications unreviewed.
The AI software analysed only a single cardiac phase, often the 70% diastolic phase, without identifying motion artefacts or flagging nondiagnostic segments. This limitation affects its capacity to replicate radiologists’ multiphase review processes. Moreover, the system currently lacks the ability to detect coronary anomalies prospectively. Multivariable analysis revealed that body mass index, sex, heart rate and CACS severity influenced AI categorisation accuracy. These findings underscore the need for continuous algorithm refinement and incorporation of diverse training data to improve performance robustness.
The evaluated on-premise AI solution for CCTA interpretation achieved substantial agreement with expert readers and demonstrated high accuracy in ruling out significant coronary artery disease. Its capacity for rapid, automated analysis makes it a valuable asset for triage and workflow support, especially in time-sensitive settings like emergency departments. While the relatively low PPV precludes its use as a standalone diagnostic tool, its integration as an aid for clinicians can enhance efficiency and patient care. Future developments should aim to improve motion artefact detection, expand multiphase analysis and address the detection of anatomical anomalies to further elevate its clinical utility.
Source: Radiology
Image Credit: iStock
