Review of AI for CMR in 2024: AI methods, CMR applications, and clinical translational outlook

Background: AI is increasingly transforming CMR in many aspects. We reviewed the most recent AI for CMR literature evidence in the last year (2024), as follow-up research to the insights gained from earlier reviews (Prieto et al., 2025; Zhang et al., 2024).

Methods: We performed PubMed searches using combinations of [“artificial intelligence”, “machine learning”, “deep learning”] and [“CMR”, “cardiac MRI”, “cardiovascular MRI”] in 2024. Review articles, case reports, and conference papers were excluded, resulting 210 original research. We categorised the literature evidence according to several new aspects, including translational stages, neural networks, patient cohorts, and use of cross-modality learning and transfer learning.

Bar and pie charts show the frequencies of translational stages, AI tasks and CMR sequences used. We used a Sankey diagram to illustrate relationships between the relevant categories in domains of the neural network types, AI tasks, and CMR sequences towards patient cohort applications.

Results: Research works in 2024 focused on three stages: technical development, clinical validation, and integration of AI in clinical workflow or research (Fig. 1). The portions of diagnosis in AI tasks, and T1/T2 mapping in CMR sequences, have increased compared to previous years.

Convolutional Neural Networks (CNN) include variations of U-Net, ResNet, DenseNet, and VGG (Fig. 2). U-Nets remain the majority, followed by transformer, generative and machine learning (ML). Segmentation remains the most common application, yet the AI tasks were balanced by a clear emerging interest in computer-aided diagnosis (classification). Within CMR sequences, applications in cine and LGE continue to dominate, consistent with their long-standing availability as the gold standard for assessing function and tissue characterization.

High quality databases are important to validate and benchmark deep learning models. There are a substantial use of public databases: UK Biobank (Sudlow et al., 2015) (n=13 studies), ACDC 2017 (Bernard et al., 2018) (n=19), M&Ms 2020 (Campello et al., 2021) (n=3), MM-WHS 2017 (Zhuang, 2019) (n=2), and Atrial Segmentation Challenge 2018 (Xiong et al., 2021) (n=2), with three new datasets introduced in 2024, for reconstruction (CMRxRecon (Wang et al., 2024)) and segmentation (RAS (Zhu et al., 2024), HVSMR-2.0 (Pace et al., 2024)). Multi-modality approaches (29%) and transfer learning (28%) are substantial, leveraging CT, echocardiography, and other pretrained models from medical and broader images.

Conclusion: The AI papers in 2024 showed increased evidence of AI translation to CMR research and clinical application through multiple translational stages and cross-disciplinary domains. By harnessing cross-modality data and diverse patient cohorts, robust and generalisable AI models would be widely anticipated. We find that the continuous collaboration between clinicians and AI developers fosters the translation from development to clinical validation and workflow integration.

Figure 1. (a) Literature review of 2024 compared with between 2019-2023. Number of original AI CMR research papers in 2024 (n= 210). 2024 papers are grouped into technical development (blue), clinical validation (green), and clinical integration (red). (b) Published AI CMR studies categorised by AI tasks. (c) Published AI CMR studies categorised by CMR sequences. Shaded regions denote the proportion of 2024 published papers. DTI: diffusion tensor imaging, ECG: electrocardiogram, ECV: extracellular volume, LGE: late gadolinium enhancement. “Whole-heart” includes coronary MR angiography. “Flow” includes 2D phase-contrast and 4D flow. “Text” includes patient characteristics, history, and CMR reports.
Figure 2. A Sankey diagram to illustrate the relationships between types of AI and its CMR applications in the original researches of 2024.