Artificial Intelligence in Cardiac MRI: Is Clinical Adoption Forthcoming?

Abstract

Artificial intelligence (AI) refers to the area of knowledge that develops computerised models to perform tasks that typically require human intelligence. These algorithms are programmed to learn and identify patterns from "training data," that can be subsequently applied to new datasets, without being explicitly programmed to do so. AI is revolutionising the field of medical imaging and in particular of Cardiovascular Magnetic Resonance (CMR) by providing deep learning solutions for image acquisition, reconstruction and analysis, ultimately supporting the clinical decision making. Numerous methods have been developed over recent years to enhance and expedite CMR data acquisition, image reconstruction, post-processing and analysis; along with the development of promising AI-based biomarkers for a wide spectrum of cardiac conditions. The exponential rise in the availability and complexity of CMR data has fostered the development of different AI models. Integration in clinical routine in a meaningful way remains a challenge. Currently, innovations in this field are still mostly presented in proof-of-concept studies with emphasis on the engineering solutions; often recruiting small patient cohorts or relying on standardised databases such as Multi-ethnic Study on atherosclerosis (MESA), UK Biobank and others. The wider incorporation of clinically valid endpoints such as symptoms, survival, need and response to treatment remains to be seen. This review briefly summarises the current principles of AI employed in CMR and explores the relevant prospective observational studies in cardiology patient cohorts. It provides an overview of clinical studies employing undersampled reconstruction techniques to speed up the scan encompassing cine imaging, whole-heart imaging, multi-parametric mapping and magnetic resonance fingerprinting along with the clinical utility of AI applications in image post-processing, and analysis. Specific focus is given to studies that have incorporated CMR-derived prediction models for prognostication in cardiac disease. It also discusses current limitations and proposes potential developments to enable multi-disciplinary collaboration for improved evidence-based medicine. AI is an extremely promising field and the timely integration of clinician's input in the ingenious technical investigator's paradigm holds promise for a bright future in the medical field.

Keywords: artificial intelligence; cardiac MRI; clinical integration; machine learning; neural network.

Conflict of interest statement

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2022 Fotaki, Puyol-Antón, Chiribiri, Botnar, Pushparajah and Prieto.

Figures

Figure 1

Simplified graph describing the three principal machine-learning methods. Supervised learning utilises hand-labelled datasets to design algorithms that predict future events, classify data into defined categories or distinguish the most relevant variables to the output. The predictive model learns through data training and improves over time. In unsupervised learning the software accomplishes the processing of raw data, finding hidden structures in datasets, without prior annotation, identifying meaningful relationships and clusters within the data. Reinforcement learning is a reward-based learning. Its foundation lies in the interactions with an environment, in which positive and negative feedback (reinforcements) contribute to the optimisation of the model.

Figure 2

Pipeline of a convolutional neural network (CNN). A CMR image functions as input to the CNN. The CNN identifies and classifies the various attributes (features) of the image for analysis in a procedure named Feature Extraction, including a stack of convolutions and pooling operations. In the convolution operation different-level features, such as edges, colour, gradient orientation are extracted from the input image. The pooling layer reduces the dimensionality of the convolved features, in order to decrease the computational requirements. The nodes in the fully-connected layer are connected directly to all nodes in the previous layer. This layer compiles the data extracted by previous layers and applies various filters to form the final output.

Figure 3

Prospective super-resolution reconstruction: coronal and coronary reformat of low-resolution acquisition (1.2 × 4.8 × 4.8 mm3) acquired in ~50 s compared to high-resolution acquisition (1.2 mm3) acquired in ~7 min. Bicubic interpolation (1.2 mm3) and proposed super-resolution reconstruction (1.2 mm3) in a patient with suspected CAD for a prospective acquired low-resolution scan (prospective cohort). Magnified image of RCA shows comparable image quality to the high-resolution acquisition in significantly shorter scan time. Küstner et al. (11). The article is published Open Access under a CC BY licence (https://creativecommons.org/licences/by/4.0/).

Figure 4

Representative image quality of the coronaries from a prospective, clinically integrated study, that utilised a residual U-Net network to facilitate super-resolution reconstruction of rapidly acquired low-resolution three-dimensional whole-heart balanced Steady State Free Precession datasets. Multi-planar reformats of the coronary artery from the respective conventional high-resolution acquisition, low-resolution acquisition, and the corresponding super-resolution reconstruction dataset. Sharpness of vascular borders is enhanced and image distortion is attenuated in the super-resolution reconstruction dataset vs. the low-resolution volume. This is particularly beneficial in the delineation of small vessels, such as the coronary arteries. Qualitative image quality analysis demonstrated no statistically significant differences between the super-resolution and the high-resolution data. Steeden et al. (13). The article is published Open Access under a CC BY licence (https://creativecommons.org/licenses/by/4.0/).

Figure 5

Examples to demonstrate the image quality and opticospatial correlation between VNE and conventional LGE images. T1 colormaps (top row) were adjusted to show the T1 signals that pair with the VNE signals. The bottom 2 rows visualise myocardial lesion regions by VNE and LGE using progressive thresholding (full width, at half, a quarter, and eighth maximum) displayed with different colours. In (A–F), high visuospatial agreement was noted between VNE and LGE. White arrows point to the lesions. Yellow arrows point to slightly different depiction of the right ventricular wall in VNE and LGE, suggesting patient movement between acquisitions. (G), An example of VNE displaying subtle changes in the distribution and quantification of the lesion clearer than LGE. LGE, late gadolinium enhancement; VNE, virtual native enhancement. Zhang et al. (16). The article is published Open Access under a CC BY licence (https://creativecommons.org/licenses/by/4.0/).

Figure 6

DL-based computation of global and segmental circumferential strain is compared to the clinician-assisted DENSE analysis. The AI-based end-systolic circumferential strain (Ecc) maps (left column), segmental (middle column) and global (right column) circumferential strain–time curves for a healthy subject (A) and a heart failure patient (B) demonstrate very close agreement with the conventional segmentation in the depicted mid-ventricular slices. Ghadimi et al. (26). The article is published Open Access under a CC BY licence (https://creativecommons.org/licenses/by/4.0/).

Figure 7

An illustrative overview of the explainable MRI concept. The user has insight in the features that influence the decision of the model.

Figure 8

Schematic representation of three proposed strategies to introduce fairness in AI algorithms. First, pre-processing modifications in the training dataset can eliminate bias before training. In each training dataset, the data are initially classified by the protected attribute(s) (such as sex, race, ethnic origin, religious and political beliefs, age, socioeconomic background and so forth). Samples are stratified to establish equitable representation of all protected groups in the training. Alternatively, alterations in the AI algorithm can train a model to overcome discrimination and optimise the performance both in the prevalent and unprivileged group(s). The third approach attempts to train distinct models for each protected group.

Figure 9

Brief chart on the framework of “clinician in the loop.” Clinicians are provided with action choices. Data labelled from clinicians contribute to the training of the network.