Image-Based Cardiac Diagnosis With Machine Learning: A Review

Abstract

Cardiac imaging plays an important role in the diagnosis of cardiovascular disease (CVD). Until now, its role has been limited to visual and quantitative assessment of cardiac structure and function. However, with the advent of big data and machine learning, new opportunities are emerging to build artificial intelligence tools that will directly assist the clinician in the diagnosis of CVDs. This paper presents a thorough review of recent works in this field and provide the reader with a detailed presentation of the machine learning methods that can be further exploited to enable more automated, precise and early diagnosis of most CVDs.

Keywords: artificial intelligence; automated diagnosis; cardiac imaging; cardiovascular disease; deep learning; machine learning; radiomics.

Figure 1

Number of publications on machine learning and cardiac imaging per year. This suggests an upward trend for future research. Light green bar represents the expected number of publications to be published late 2019.

Figure 2

Pipeline for building image-based machine learning models.

Figure 3

Input variables type distribution in reviewed literature. As seen in the pie chart, conventional indices are the predominant features for training ML models, followed by radiomics and deep learning techniques.

Figure 4

Summary of common input and output variables for image-based diagnosis ML algorithms. Different cardiac imaging input features such as raw data, conventional indices extracted from a ROI or radiomics (delineation of cardiac anatomy is required for the last two cases) and desired output. Both structures shape the most basic requirement for a ML cardiac imaging application, data.

Figure 5

Machine Learning technique distribution.

Figure 6

Selected machine learning techniques. (A) Logistic Regression is used to model the probability of a binary outcome. In the figure, Y axis represents the probability while X axis is the continuous input variable. Notice that small changes in X produce large variations of the final probability Y, mainly in the central part of the plot where the uncertainty of the model is larger. This model can be extended to a multi-class problems. (B) Support Vector Machine models are able to transform a non-linear boundary to a linear one using the kernel trick. During the training process, the distance between classes to the final selected boundary is maximized. (C) Random Forest is a technique that combines Decision Trees for reducing the uncertainty in the final prediction. It is based in a recursive binary splitting strategy where upper nodes are intended to be the most discriminative ones and subsequent branching is applied to less relevant variables. (D) Clustering is a technique with capability to find subgroups (clusters) along data. There are different cluster techniques, some need a prior number of clusters (kMeans), some of them can be used with output information (kNN), and others are fully unsupervised (meanShift). (E) Artificial neural networks are able to model complex non-linear relations between input variables and outcomes by propagating structured data (green nodes—input variables), e.g., radiomics, through hidden layers (blue nodes) to obtain an output (orange nodes). (F) Convolutional neural networks are the backbone of Deep Learning applications. They comprise input and output layers separated by multiple hidden layers. Their ability to hierarchically propagate imaging information and extract data-driven features implies automatic detection of relevant cardiac imaging biomarkers within the intermediate layers.

Figure 7

Distribution of image-based diagnostic application using machine learning (A) per disease, (B) per modality.

Figure 8

Factors involving robustness and reproducibility of quantitative imaging features.