Present and Future Innovations in AI and Cardiac MRI

Abstract

Cardiac MRI is used to diagnose and treat patients with a multitude of cardiovascular diseases. Despite the growth of clinical cardiac MRI, complicated image prescriptions and long acquisition protocols limit the specialty and restrain its impact on the practice of medicine. Artificial intelligence (AI)-the ability to mimic human intelligence in learning and performing tasks-will impact nearly all aspects of MRI. Deep learning (DL) primarily uses an artificial neural network to learn a specific task from example data sets. Self-driving scanners are increasingly available, where AI automatically controls cardiac image prescriptions. These scanners offer faster image collection with higher spatial and temporal resolution, eliminating the need for cardiac triggering or breath holding. In the future, fully automated inline image analysis will most likely provide all contour drawings and initial measurements to the reader. Advanced analysis using radiomic or DL features may provide new insights and information not typically extracted in the current analysis workflow. AI may further help integrate these features with clinical, genetic, wearable-device, and "omics" data to improve patient outcomes. This article presents an overview of AI and its application in cardiac MRI, including in image acquisition, reconstruction, and processing, and opportunities for more personalized cardiovascular care through extraction of novel imaging markers.

Graphical abstract

Figure 1:

Diagram shows the progress, opportunities, and current challenges in the use of artificial intelligence in cardiac MRI.

Figure 2:

Diagram of supervised learning strategy for an artificial intelligence model. An architectural layer performs computational operations on data passing through it, with skip connections linking nonsequential layers. A forward pass transforms input into output. The output is compared with an ideal or correct ground truth. The task of the model (eg, image reconstruction, automated analysis, or clinical diagnosis) dictates the form of “ground truth,” which could be high-quality images, expert annotations, or patient records. The loss function quantifies the discrepancy between the output and the ground truth, converting the disagreement into an error signal. The backward pass propagates the error back through the model. This is accomplished by calculating gradients, which represent the direction and rate of change of the error, and selecting an optimizer, which determines how the architecture parameters are updated based on these gradients. This process reduces the error over time. An epoch is one full pass over the training data set, with the architecture trained across many epochs until error convergence is achieved.

Figure 3:

Diagram of common learning strategies for artificial intelligence models. Supervised learning trains models on input-output relationships using “ground truth” as a reference to guide the learning process. Semisupervised learning, with partial or imperfect ground truth, is effective where only partial supervision is feasible. Transfer learning refines a preexisting model using related data or tasks, saving time and computational resources compared with starting from scratch. Ensemble learning combines outputs from multiple uniquely trained models, reducing error risks and enhancing model robustness for more reliable, accurate predictions.

Figure 4:

Diagram of accelerated cardiac MRI reconstruction techniques. Accelerated imaging shortens scan time by collecting only partial k-space data instead of a complete set. The traditional reconstruction uses inverse fast Fourier transform (ifft) for k-space-to-image transformation, causing aliasing artifacts in accelerated imaging. K-space-agnostic artificial intelligence (AI) methods transform aliased images (input) into clean, high-quality images (output) without detailed knowledge of the acquisition process. Conversely, k-space-aware AI methods use information of the acquisition process as prior knowledge, integrating image-to-image models with alternating k-space data consistency (DC) steps to better handle incomplete k-space data. Alternatively, AI models can convert partial k-space inputs into a complete set, aiding in generating cleaner, higher-quality images. fft = fast Fourier transform.

Figure 5:

Diagram of artificial intelligence (AI)–based image reconstruction (recon) and enhancement. The sampling trajectory dictates the k-space data collection path; a Cartesian trajectory is a grid-like path. Reconstruction latency is the delay from data collection to viewing the image. K-space data are collected over multiple heartbeats via electrocardiogram (ECG)–segmented imaging, usually using a fully sampled Cartesian trajectory, which aids in functional imaging and minimizes motion artifacts at the cost of longer scan time. Accelerated imaging, employing undersampled Cartesian, radial, or spiral sampling, reduces scan time by acquiring partial k-space data. However, utilizing a simple inverse fast Fourier transform (iFFT) for frequency (ie, k-space)–to-image domain conversion causes artifacts like incoherence, streaking, and swirling in images. Various AI reconstruction models mitigate these artifacts, improving image clarity and quality. Another method to reduce scan time is truncating the k-space data, which yields lower-resolution images without aliasing artifacts. AI enhancement is used in this context to restore the lost spatial resolution.

Figure 6:

Artificial intelligence (AI)–based image reconstruction. Free-breathing electrocardiogram-free real-time cardiac MRI scans collected with highly accelerated imaging sequences have substantial aliasing artifacts. Convolutional U-Net models are trained using paired high-quality and aliased images to remove such artifacts. (A) A single U-Net model can remove artifacts from balanced steady-state free precession real-time cine images, as shown in these diastolic and systolic cardiac phase MRI scans in the short-axis plane. (B) Gradient-recalled echo real-time phase-contrast flow imaging acquires velocity-compensated and velocity-encoded images, as shown in these aortic outflow tract plane MRI scans. Two separate U-Net models were used to remove the artifacts while preserving the phase-contrast information.

Figure 7:

Artificial intelligence (AI)–based resolution enhancement. Collecting balanced steady-state free precession cardiac MRI scans with lower resolution reduces scan time since fewer data are acquired. Acceleration reduces the breath-hold time needed in breath-hold electrocardiogram (ECG)–gated sequences or enables free-breathing real-time imaging with high temporal resolution. (A) Cine images in the short-axis plane show enhancement of spatial resolution using a generative adversarial network. (B) Tagging images in the short-axis plane show enhancement of spatial resolution using a pretrained model based on cine images.

Figure 8:

Artificial intelligence (AI)–based resolution enhancement. A pretrained AI-based resolution enhancement model based on balanced steady-state free precession cardiac MRI cine data can be used to accelerate inversion recovery late gadolinium enhancement imaging, as shown in images in the short-axis plane. For instance, current existing sequences require a 16-second breath hold (BH) for imaging, with one breath hold for each section. The imaging time can be reduced to 10 seconds or even 6 seconds at the expense of diminished spatial resolution. The AI model is used to enhance the spatial resolution in images with shorter imaging time.

Figure 9:

Diagram shows artificial intelligence (AI)–enabled cardiac MRI analysis, interpretation, and reporting. In cardiac MRI, images are collected to characterize function (ie, cine and tagging images), flow, tissue properties and fibrosis (ie, T1 and T2 maps and T1-weighted, T2-weighted, and extracellular volume [ECV] images), and scarring (ie, late gadolinium enhancement [LGE] images). AI can substantially impact the analysis workflow after imaging. The clinical reading includes quantification (eg, function on cine images) and clinical interpretation of individual sequences (eg, presence and location of scarring on LGE images). AI can assist in both the analysis and interpretation of clinical readings and can change the role of imagers to oversight.

Figure 10:

Automated segmentation using convolutional U-Net convolutional neural networks. Automated segmentation based on U-Net architectures has been demonstrated across many cardiac MRI sequences, including cine, T1 and T2 mapping, late gadolinium enhancement (LGE), and perfusion images.

Figure 11:

The impact of artificial intelligence (AI)–enabled cardiac MRI on the patient care pipeline. The diagram shows how AI impacts information extraction and flow at different steps. (A) AI facilitates and improves the analysis and interpretation of images by extracting standard cardiac MRI parameters. This reduces the analysis burden and improves the accuracy, precision, and reproducibility of analysis and interpretation. AI may also provide a new paradigm for gaining insights into cardiac MRI scans by extracting radiomic or deep imaging signatures of cardiac disease not currently being extracted. (B) AI enables the efficient combination of clinical, imaging, wearable device, biomarker, genetics, and “omics” data to provide clinically actionable information to improve patient care. LGE = late gadolinium enhancement.