Artificial intelligence in cardiac imaging: where we are and what we want

‘Artificial Intelligence is the new electricity’ Andrew Ng Artificial intelligence (AI) is spreading to every aspect of cardiac imaging, including study indication, protocol selection, image acquisition and reconstruction, data post-processing, analysis, quantification, and to the holy grail of medicine: patient-specific risk prediction and prognostication. Currently, many of these applications are making their way from the research domain to clinical practice. More than 300 AI-enabled devices have been Food and Drug Administration (FDA) approved in the combined radiology and cardiovascular domains, and this number is continuously growing. Among the many advantages of the use of AI are the possibilities to increase efficiency, reduce human error and reader variability, and to increase the accuracy of diagnosis and prognosis. Echocardiography, coronary computed tomography angiography (CCTA), and cardiac magnetic resonance imaging (CMR) imaging can significantly benefit from AI-based solutions, making these techniques easier to deploy in clinical setting by improving reproducibility in obtaining quantitative morphological and functional information, all of which can improve patient care and outcomes. In the field of imaging acquisition and reconstruction, convolutional neural network-based solutions are helping to drastically reduce radiation and contrast dose for CCTA imaging, while focusing on faster reconstructions and acquisition in CMR imaging thanks to advanced signal under-sampling and real-time acquisitions, improving the overall image quality. For cardiac CT imaging, many algorithms for fast and accurate evaluation of the coronary tree are under development or already available for clinical deployment (Figure 1). Coronary artery calcium (CAC) scoring can be easily automated and is one of the first tasks that has been tackled with AI. AI-based CAC scoring has excellent agreement with human readers in a fraction of the time and as a result is being employed clinically (Figure 1A).1 The Coronary Artery Disease-Reporting and Data System (CAD-RADS) 2.02 is recommended for CCTA to standardize reporting, communication and decision-making. It is more complex than CAC scoring and therefore can be time consuming and more prone to inter-reader variability, even among expert readers. AI-powered software to provide interpretation of stenosis and plaque analysis, which is needed for automated CAD-RADS reading, are under development and have demonstrated improved inter-reader variability compared to experienced readers (Figure 1B and C). AI algorithms for Agatston score, coronary stenosis, and CT-FFR assessment. (A) Clinically available AI algorithms for the detection of coronary calcium and Agatston score quantification (Syngo.via VB50 Siemens Healthineers); (B and C) Pre-clinical prototype for automated plaque detection, stenosis severity quantification, and CAD-RADS classification (AI-Heart, Siemens Healthineers); (D) Research software for coronary MLCT-FFR assessment (CT-FFR prototype, Siemens Healthineers). Currently, computational fluid dynamics (CFD) computed tomography fractional flow reserve (CT-FFRCFD, Heartflow Inc.) is the only FDA approved clinical solution for CT-FFR assessment. However, machine learning-based CT-FFR (CT-FFRML) has been developed for rapid in-house CT-FFR assessment for research application (Figure 1D). In a large multicentre clinical trial,3 it demonstrated a similar accuracy of AI-based CT-FFR compared with CFD-based CT-FFR, which has strengthened interest in developing this AI-based technology for clinical implementation. In the future, these AI algorithms for CCTA will likely incorporate identification of vulnerable plaque features and quantification of plaque burden that will further improve cardiovascular disease risk assessment, as both have shown value in identifying patients at risk for adverse outcomes.4,5 These AI algorithms may also be used to evaluate disease progression and effectiveness of lipid-lowering therapies for plaque progression (Figure 2A and B). AI algorithms for advanced CCTA plaque analysis and CMR ventricular volumes assessment. (A) AI-based coronary plaque burden quantification and identification of high-risk plaque features, indicated by the yellow box and arrows (Cleerly Inc.); (B) Coronary plaque burden quantification can be utilized to monitor disease progression or response to therapy (Elucid Bioimaging Inc.); (C) Automated cardiac chambers segmentation for rapid volumes and function assessment (Circle Cardiovascular Imaging Inc.). CMR provides highly quantifiable data on cardiac function, valvular disease, perfusion, and blood flow. Tissue characterization is an important feature of CMR and is mostly focused on T1 or T2 images and late gadolinium enhancement. Currently, CMR requires manual segmentation of cine images for the accurate assessment of left and right ventricles volumes and function, a time-consuming process prone to inter-reader variability. The use of AI for segmentation can increase workflow effienciency, reduce variability in readings, and automate LV function quantification6 (Figure 2C). Additionally, with the assistance of FDA approved AI software, transthoracic echocardiographic images can now be obtained by untrained operators, and similar algorithms for automated fast CMR acquisition are under development, which may facilitate care in settings with limited resources. What we want, what we need, and what we receive are often misaligned, especially in medicine. Regarding cardiac imaging, what we need and can reasonably achieve short-term are automated, fast, and reproducible AI-based integrated solutions for anatomical and functional evaluation of the coronary system and myocardium, seamlessly integrated within our clinical workflow. What we should aim for is multimodality integration, with data from echocardiography, CT, CMR, SPECT, and PET, and so on automatically incorporated in a patient-tailored unique cardiac model that can serve as a digital twin to our individual patients. Ideally, data from multimodality imaging sources need to be automatically integrated with clinical data from the electronic medical record, to provide real-time accurate patient-specific comprehensive representation of the heart, including the myocardium, valves, coronaries, disease severity and functional impact to guide treatment and predict the outcome of interest. With the rise of new imaging analysis techniques supported by AI such as radiomics, the value of all these newly developed biomarkers as well as new risk stratification and prognostic models need to be assessed to allow translation to clinical practice. Because of its capabilities in analyzing large amounts of data and enabling the detection of complex relationships, AI algorithms are an ideal way to create predictive models that leverage cardiac imaging. To reach these goals, we need large, heterogenous datasets with multi-modality imaging to overcome the systematic underrepresentation of women and minority populations in clinical trial and avoid any form of bias, a well-known problem in medicine and AI, which has attracted criticism. Finally, with AI already gaining a place in the clinical workflow, it is necessary to prove the value of AI for patient care, to justify the cost of these new algorithms and drive new reimbursement or payment models. For this to happen on a larger scale, cost-effectiveness studies of AI technology are essential. Artificial intelligence is changing the medical sector and is improving cardiac imaging. AI-based cardiac imaging applications can play a significant role in the identification, quantification, and characterization of cardiac disease, that rapidly and accurately provides physicians with a more comprehensive description of a patient’s cardiovascular health allowing for better decision making to improve patient outcomes. We need to embrace and guide this revolution to improve the quality of healthcare, minimize disparities in the access to medical services and reduce the cost of medical treatments. C.N.D.C. receives funding from Siemens Healthineers, and he is a consultant for Covanos Inc. M.v.A. receives research funding from Siemens Healthineers. The others have no disclosures to report. No new data were generated or analysed in support of this research.