Artificial Intelligence in Cardiovascular Imaging- The Basics

Contributed by Dr Navni Garg and Dr Vishnu Vardhan Ravilla

Artificial intelligence (AI) denotes to a set of methods where a computational program can perform tasks that are characteristic of human intelligence. Typically such tasks include, pattern recognition and identification, planning, understanding language, recognizing objects and sounds, and problem-solving. Humans make decisions using following faculties, namely perception, diligence and cognition. AI can support this decision making by assisting them in avoiding perceptual errors. Machine Learning (ML) is a branch of AI where computers learn from human decisions on past data points so that they can take decisions on new data points based on what it has learnt. However, Deep Learning (DL) structures algorithms in layers to create an ‘artificial neural network’ which has the capability of learning and making intelligent decisions on its own. DL models can perform automatic feature extraction from raw data. The computer undergoes multiple stages of learning and aims at forming a complex higher level of features from the composition of simple low-level features.

For developing an algorithm, a large amount of heterogeneous data is collected and segmented into training and validation datasets. The algorithm is trained on the training dataset to find different patterns related to the problem we are trying to solve. The trained model is then validated on the validation dataset. Increasing the amount and quality of input data improves the performance of AI-based models. The trained model is then tested on independently acquired data for its accuracy and precision.

In healthcare, large amounts of collected data including medical health records, age and gender distribution of diseases, patient results, outcome data, genomic data, biomarker analysis, and image-derived information, to name a few, can be utilized for machine learning. The collected data segmented into data points like pixel density, signal intensity or brightness, vector of motion or measurement from images can be utilized either in supervised or an unsupervised manner. In medical imaging, the machine can be trained to decipher Echocardiography, CT Scans, X- Rays and MRI Scans for simple patterns as well as for comprehensive interpretation.

AI in Cardiovascular Imaging:
There has been a global increase in cardiac imaging investigations over the years because of broader availability and acceptability. Use of medical AI can decrease the workload of radiologists by performing some mundane, time-consuming tasks as well as enhance their skill and decision making.

Chest X rays (CXR): Conventional radiographic features of congestive heart failure include cardiomegaly, pulmonary venous congestion, septal lines, airspace opacification, and pleural effusions. There are AI-based models trained to predict this appearance of heart failure on CXR (1).

Echocardiography (ECHO): It is widely available and utilized. It is often the first imaging modality used in cardiology. However, one of the drawbacks of ECHO is that it is operator dependent. AI models can reduce its user dependence and achieve more standardized results. ML-based models can automatically identify and measure the left ventricular (LV) wall thickness during the acquisition of echocardiographic images. These models have shown comparable results to 3D echocardiographic methods and cardiac MRI. However, in patients with small ventricular cavities, LV wall may not be optimally visualized (2). AI trained models can also help in the classification of echocardiographic views, determination of left ventricular ejection fraction and longitudinal strain. Automated segmentation, analysis of left and right ventricle contours, identification of MI and its severity, characterization of the phenotype of heart failure with preserved ejection fraction and automated calculation of volumetric parameters, are few areas where AI finds its application. AI hopes to increase the efficiency and decrease the workload of cardiologists and echocardiographers and assist them in better diagnosis.

Cardiac CT: AI can aid in the acquisition of a low dose CT scan. It can also help determine calcium score from regular coronary CT angiography (CTA) and chest scans without having to perform a separate calcium score CT. AI has made cardiac post-processing less cumbersome and reasonably accurate, e.g., Automated segmentation of the left ventricle, detection and characterization of coronary plaques, degree of stenoses and differentiation between plaques with and without napkin ring sign (NRS). Significant coronary stenosis is defined as a fractional flow reserve (FFR) <0.8 determined during invasive coronary angiography (ICA). Automated quantification FFR from CT coronary scan data, is another critical and recent application. For this, geometric features of the coronary anatomy, physiological characteristics, quantitative plaque measurements, features calculated from different spatially connected clusters of heart segmentation and CT perfusion features have been used to train the AI models. AI-based software can also automatically identify anatomical structures around and inside the heart and colour code them. Such post-processing applications make calculations more accurate for radiologists, post-processing time-efficient for technologists and interpretation easier for referring physicians. Additionally, AI-based models can identify patients at risk of cardiac events. The role of AI in cardiac CT is expanding, and with time and supervision from experts, the performance of AI models will improve, and other new features may be available.

Cardiac MRI: Cardiac MRI is an investigation that involves acquiring images of a moving structure, the heart, which is caged by the lungs and diaphragm that are all moving at the same time, although independent.  Therefore, the inherent relative disadvantage in comparison to MRI scans of other parts of the body, include long scanning time and time-consuming post-processing techniques. The quality of cardiac MRI images is patient, user, scanner and vendor dependent leading to inter and intra-patient variability posing a diagnostic challenge to use of AI in cardiac imaging. Thus, the primary use of AI-based models is to reduce the time of the acquisition and therefore make it more patient-friendly, reducing inter-technologist variability in planning, and reducing post-processing time for technologists and radiologists.  Hastening cardiac MRI scan would reduce the breath-hold related factors that lead to the degradation of image quality. AI-driven software help acquire cardiac ischemia exam in less than 15 minutes. It can also detect artefacts and prompt the operator to reacquire images if the image quality is suboptimal. Some of the post-processing software currently available includes automatic segmentation of anatomy, semi-quantitative perfusion analysis, quantitative delayed enhancement analysis and functional quantification measurements in cardiac MRI. Automatic segmentation of enhancement on late gadolinium enhancement imaging for differentiation between patients with myocardial infarction and control subjects have been studied using AI-based models.  Efforts to develop ML-based models that are capable of predicting cardiovascular risk when supplied with clinical information and data from CMR, cardiac CT and echocardiography are on.
Potential role of AI:
  • Streamline work processes in healthcare and imaging
  • Hasten image acquisition and reconstruction
  • Allow image classification
  • Automate the analysis of variables by performing time-consuming calculations
  • Assist image interpretation and reporting
  • Guide referring physicians with diagnosis and prognostication based on clinical and radiological parameters
  • Auto-triage patients who need immediate care in emergency departments

AI-based software is bound to have lower error rates, greater accuracy, incredible precision and speed as compared to humans only if appropriately coded. Setting up AI-based machines and developing AI-based models entails enormous cost. Also, AI-based programs require regular maintenance and up-gradation.

AI has not achieved its full potential in healthcare due to lack of expertly labelled data, privacy-concerns, legal issues for the use of data and resistance from physicians and imagers. To begin with, AI is only as good as the expert-mind, and the quality and quantity of the data sets used to develop the AI model. It gets better with time as more heterogeneous data sets fed into these models. An expert would always be required to check the precision of AI-based models at the user or clinical end due to the real-life case heterogeneity. AI is not a threat to the medical fraternity. Instead, it will make imaging and clinical decisions time-efficient, reduce human error and in the process also enhance skills and decision-making capacity. Clinically robust AI algorithms can reduce the stress on medical professionals by performing tasks that can be automated so that they can focus on tasks that require human intelligence. But before clinical deployment, multicenter trials are essential for validation of generalizability of AI models to varied populations.

  1. Seah JCY, Tang JSN, Kitchen A, Gaillard F, Dixon AF. Chest radiographs in congestive heart failure: visualizing neural network learning. Radiology2019;290(2):514–522.
  2. Tamborini G, Piazzese C, Lang RM, et al. Feasibility and Accuracy of automated software for transthoracic three-dimensional left ventricular volume and function analysis: comparisons with two-dimensional echocardiography, three-dimensional transthoracic manual method, and cardiac magnetic resonance imaging. J Am Soc Echocardiogr. 2017;30(11):1049–58.
• Medical health records
• Age and gender distribution of diseases
• Patient results
• Known risk factors for disease
• Outcome data
• Genomic data
• Biomarker analysis
• Edge, shape, margins
• Colour and contrast
• Signal intensity
• Density
• Vector of motion
• Brightness
• Voxel measures
• Attenuation values




• Guide the operator performing ECHO
• Obtain low dose CT
• Acquire rapid cardiac MRI
• Detect motion artefacts in cardiac MRI

• Classification of cardiac views
• Automatic segmentation, analysis of left and right ventricle contours
• Identification of MI and its severity
• Determination of left ventricular ejection fraction and longitudinal strain
• Characterization of the phenotype of heart failure with a preserved ejection fraction
• LV wall thickness measurements on echocardiography
• Automated calculation of CT calcium score
• Detection and characterization of coronary plaques on CTA
• Derive FFR from coronary CT angiography
• Automatic segmentation of anatomy
• Semi-quantitative perfusion analysis
• Quantitative delayed enhancement analysis
• Functional assessment

• Predict heart failure on CXR
• Identify patients at increased risk of cardiac events