AI-Powered CT Perfusion Software Shows Promise in Detecting Myocardial Ischemia
In a significant advancement for cardiovascular imaging, a newly developed artificial intelligence (AI)-driven software tool is streamlining the analysis of dynamic CT myocardial perfusion (CTP), offering clinicians a faster, more accurate method to detect myocardial ischemia. The study, led by Zhao Runtao and colleagues from the Chinese PLA General Hospital and published in the International Journal of Medical Radiology, introduces Myocardiac Kit (MK), a semi-automated analysis platform designed to overcome longstanding barriers in the clinical adoption of dynamic CTP.
For years, dynamic CTP has held promise as a non-invasive method to assess myocardial blood flow and identify regions of compromised perfusion—key indicators of ischemic heart disease. Unlike static CTP, which provides only relative or semi-quantitative data, dynamic CTP captures the full time course of contrast enhancement, enabling true quantitative measurement of myocardial blood flow (MBF). This capability allows for a more physiologically grounded assessment of coronary artery disease, potentially reducing the need for invasive procedures such as coronary angiography and fractional flow reserve (FFR) measurements.
However, despite its diagnostic potential, dynamic CTP has struggled to gain widespread clinical traction. The primary obstacles have been the complexity of data analysis, the time-intensive nature of manual processing, lack of standardized protocols, and variability in diagnostic thresholds across studies. These challenges have limited reproducibility and increased the dependency on operator expertise, undermining confidence in the consistency of results.
The research team, comprising experts from the Department of Cardiology at the Second Medical Center of the Chinese PLA General Hospital, the Sixth Medical Center of the PLA General Hospital, and GE Healthcare China, set out to address these limitations by developing MK—a software solution that integrates AI algorithms to automate and standardize the processing of dynamic CTP data. The goal was not only to improve diagnostic accuracy but also to enhance workflow efficiency, making the technology more accessible in routine clinical settings.
The study enrolled 87 patients—67 men and 20 women—with a mean age of 60.98 years—who were suspected of having coronary artery disease. All participants underwent a combined protocol involving dynamic CTP and coronary CT angiography (CCTA), followed by invasive coronary angiography (ICA) with FFR measurement within one week. This design allowed for direct comparison between the non-invasive imaging findings and the gold-standard invasive assessment.
Dynamic CTP scans were performed using a dual-source CT scanner (Siemens Somatom Definition Flash) under adenosine-induced stress, a pharmacological agent that mimics the effects of exercise on coronary blood flow. The imaging protocol prioritized CTP, with a 28-second scan capturing contrast dynamics in the left ventricle. This was followed by a CCTA scan after heart rate stabilization, enabling the simultaneous evaluation of coronary anatomy and myocardial perfusion—a “one-stop-shop” approach that is increasingly favored in cardiac imaging.
The raw CTP data were processed using the MK software, which employs a deep learning-based convolutional neural network to automatically segment the left ventricular myocardium. The software identifies the optimal phase for perfusion analysis and divides the myocardium into 17 standardized segments, following the American Heart Association model. Users manually define the long axis of the heart and adjust segmentation if necessary, but the bulk of the analysis—including the generation of time-attenuation curves (TACs), calculation of arterial input function (AIF), and derivation of quantitative parameters—is fully automated.
Key parameters derived by MK include myocardial blood flow (MBF), myocardial blood volume (MBV), mean transit time (MTT), and time to maximum (Tmax). The software also computes semi-quantitative metrics such as the area under the curve (AUC), maximum slope, peak contrast concentration, and time to peak. These values are visualized in polar maps (bull’s-eye plots), facilitating intuitive interpretation by clinicians.
Two experienced cardiologists, each with over seven years of experience in coronary CT reconstruction, independently analyzed the data using MK and a standard CCTA workstation. The time required for data processing was recorded, and inter-observer agreement was assessed using the intra-class correlation coefficient (ICC). Diagnostic performance was evaluated against the reference standard of ICA stenosis ≥90% or FFR ≤0.80, both widely accepted criteria for defining hemodynamically significant ischemia.
The results demonstrated strong consistency between the two observers. At the myocardial segment level (n = 1,479 segments), ICC values for MBF, MBV, and MTT ranged from 0.66 to 0.76, indicating good agreement. Semi-quantitative parameters such as AUC, maximum slope, and time to peak showed even higher consistency, with ICCs exceeding 0.80. At the vascular level (n = 261 coronary arteries), similar patterns were observed, with MBF showing an ICC of 0.69, again reflecting acceptable inter-observer reliability.
More importantly, the software effectively differentiated ischemic from non-ischemic myocardium. In ischemic segments, defined by the reference standard, the mean MBF was 123.14 ± 41.83 mL/100 mL/min, significantly lower than the 147.47 ± 43.98 mL/100 mL/min observed in non-ischemic segments (P < 0.05). Similarly, myocardial segments supplied by ischemic coronary arteries exhibited lower MBF (124.34 ± 42.86 mL/100 mL/min) compared to those supplied by non-ischemic vessels (148.68 ± 44.49 mL/100 mL/min), further validating the physiological relevance of the measurements.
To determine the optimal diagnostic threshold, the researchers used the Youden index to identify the MBF cutoff that best discriminated between ischemic and non-ischemic territories. At the vascular level, the optimal cutoff was found to be 115.0 mL/100 mL/min. Using this threshold, MBF alone achieved an area under the ROC curve (AUC) of 0.76, with a sensitivity of 79% and specificity of 71%. In contrast, relying solely on CCTA-defined diameter stenosis ≥50% yielded a lower AUC of 0.62, with high sensitivity (90%) but poor specificity (42%), reflecting the well-known limitation of anatomical imaging in overestimating functional significance.
The most compelling finding emerged when MBF and CCTA stenosis were combined. This integrated approach achieved an AUC of 0.91 (95% CI: 0.87–0.95), with sensitivity of 84%, specificity of 80%, and overall accuracy of 82%. This synergistic effect underscores the value of combining anatomical and functional data—a paradigm increasingly recognized in precision cardiology.
From a practical standpoint, the efficiency gains offered by MK were substantial. The average time for data analysis using the software was just 10.51 ± 1.95 minutes, a significant reduction compared to traditional manual methods, which often exceed 40 minutes. Even among existing semi-automated tools, which typically require 15 minutes or more, MK demonstrates superior speed, likely due to its optimized AI-driven segmentation and streamlined workflow.
Radiation exposure, another critical consideration in cardiac CT, was also favorable. The combined CTP and CCTA protocol resulted in an effective radiation dose of 4.83 ± 1.84 mSv, well within acceptable limits and comparable to or lower than many standalone cardiac CT exams. This was achieved through the use of low-kilovoltage scanning (70 kV for CTP) and ECG-tailored tube current modulation, reflecting a commitment to dose optimization.
The technical foundation of MK lies in its use of a physiological model known as the single-compartment or “single circulation” model. This model assumes that contrast agent enters the myocardium via arteries, remains within the capillary bed during the first pass, and exits through veins. By applying the principle of mass conservation and contrast dilution theory, the software calculates the residue function R(t), which reflects the tissue’s ability to retain contrast. Using deconvolution algorithms, MK derives MBF from the relationship between the arterial input function and the tissue concentration curve.
What sets MK apart is its integration of deep learning for myocardial segmentation. The software utilizes a Dense V-Network architecture—a type of convolutional neural network known for its efficiency in medical image segmentation. This allows for rapid and accurate delineation of the left ventricular myocardium, minimizing the need for manual correction. While some user input is still required (e.g., defining the long axis), the automation level significantly reduces subjectivity and variability.
The implications of this study are far-reaching. By demonstrating high reproducibility, diagnostic accuracy, and operational efficiency, MK addresses the core barriers that have hindered the adoption of dynamic CTP. The ability to perform a comprehensive cardiac assessment—evaluating both coronary anatomy and myocardial perfusion—in a single, non-invasive exam could transform the diagnostic pathway for patients with suspected coronary artery disease.
Currently, many patients undergo CCTA first. If significant stenosis is found, they may be referred for invasive angiography with FFR measurement to determine whether revascularization is necessary. This two-step process is not only costly and time-consuming but also exposes patients to procedural risks. A reliable non-invasive alternative that combines anatomical and functional assessment could reduce the number of unnecessary invasive procedures, improve patient outcomes, and optimize resource utilization.
Moreover, the use of AI in this context exemplifies the shift toward data-driven, precision medicine. Rather than relying on subjective visual assessment, MK provides objective, quantitative metrics that can be tracked over time, enabling more nuanced monitoring of disease progression and response to therapy. This is particularly valuable in patients with microvascular dysfunction, where coronary arteries appear normal on angiography but perfusion is impaired—a condition that is increasingly recognized but difficult to diagnose.
The study does have limitations. It was conducted at a single center with a relatively small sample size, and the patient population was selected based on clinical suspicion of coronary disease, which may limit generalizability. Additionally, while the software reduces manual input, it does not eliminate it entirely—user-defined axis placement and occasional segmentation corrections could introduce variability. Future iterations may aim for full automation, including automatic axis detection and subendocardial/subepicardial layer analysis, which could further enhance precision.
Another area for exploration is the application of MK in different clinical scenarios, such as acute chest pain, post-revascularization assessment, or evaluation of cardiomyopathies. The software’s ability to quantify MBF could also be leveraged in research settings to study the effects of novel therapies on myocardial perfusion.
Nonetheless, the current findings represent a meaningful step forward. As cardiovascular imaging continues to evolve, the integration of AI into clinical workflows is no longer a futuristic concept but a present-day reality. Tools like MK are not meant to replace radiologists or cardiologists but to augment their expertise, allowing them to focus on interpretation and patient care rather than tedious data processing.
In conclusion, the development and validation of the Myocardiac Kit software mark a pivotal moment in the evolution of cardiac CT. By harnessing the power of artificial intelligence to deliver fast, consistent, and accurate perfusion analysis, the tool has the potential to make dynamic CTP a routine component of cardiovascular evaluation. As the technology matures and undergoes broader validation, it may well become a cornerstone of non-invasive ischemia detection, improving diagnostic confidence and patient outcomes across healthcare systems.
Zhao Runtao, Dou Guanhua, Wang Fan, Wang Kai, Shan Dongkai, Wang Sicong, Yang Junjie; Chinese PLA General Hospital; International Journal of Medical Radiology; DOI: 10.19300/j.2021.L19138