CineMA: A Foundation Model for Cine Cardiac MRI

Abstract: Cardiac magnetic resonance (CMR) is a key investigation in clinical cardiovascular medicine and has been used extensively in population research. However, extracting clinically important measurements such as ejection fraction for diagnosing cardiovascular diseases remains time-consuming and subjective. We developed CineMA, a foundation AI model automating these tasks with limited labels. CineMA is a self-supervised autoencoder model trained on 74,916 cine CMR studies to reconstruct images from masked inputs. After fine-tuning, it was evaluated across eight datasets on 23 tasks from four categories: ventricle and myocardium segmentation, left and right ventricle ejection fraction calculation, disease detection and classification, and landmark localisation. CineMA is the first foundation model for cine CMR to match or outperform convolutional neural networks (CNNs). CineMA demonstrated greater label efficiency than CNNs, achieving comparable or better performance with fewer annotations. This reduces the burden of clinician labelling and supports replacing task-specific training with fine-tuning foundation models in future cardiac imaging applications. Models and code for pre-training and fine-tuning are available at https://github.com/mathpluscode/CineMA, democratising access to high-performance models that otherwise require substantial computational resources, promoting reproducibility and accelerating clinical translation.

CineMA: A Foundation Model for Cine Cardiac MRI

The study presents CineMA, a foundation model aimed at enhancing the automation of cine cardiac magnetic resonance imaging (CMR) tasks such as ejection fraction (EF) estimation, ventricle and myocardium segmentation, disease detection and classification, and landmark localisation. The model leverages self-supervised learning without requiring manual annotations upfront, a key factor in minimizing the burden traditionally placed on clinicians for image labeling.

Cardiovascular diseases (CVDs), a significant global health challenge, necessitate accurate and efficient diagnostic tools. Cine CMR, known for its high tissue contrast and non-invasive nature, plays a crucial role in the assessment of cardiac size, mass, and function. Despite its importance, deriving key biomarkers from cine CMR, such as EF, remains labor-intensive and subjective, necessitating the expertise of clinicians to avoid variability in assessment outcomes.

Methodology and Findings

CineMA employs a self-supervised masked autoencoder framework trained on a vast collection of 74,916 cine CMR studies from the UK Biobank, encompassing multiple views such as SAX and LAX. CineMA's design facilitates tasks across eight datasets covering 23 tasks, divided into four categories—segmentation, EF estimation, disease detection and classification, and landmark localisation.

Notably, CineMA demonstrated remarkable accuracy in EF estimation from segmentation with a mean absolute error (MAE) of 3.33\% for LVEF and 5.49\% for RVEF, outperforming CNN counterparts with errors of 3.63\% and 6.28\%, respectively. The model showed enhanced performance with reduced error margins on datasets not included during pre-training, further underscoring its robustness in zero-shot learning scenarios.

Moreover, CineMA exhibited high label efficiency, performing comparably with only 50\% of labels, thus significantly reducing annotation requirements. In disease detection tasks, CineMA achieved a specificity of 60.06\% versus 31.55\% by CNN, while maintaining high sensitivity at 90.20\%.

Implications and Future Directions

The foundation model approach adopted here signifies a shift from task-specific models to adaptable general-purpose models. CineMA’s robust performance across multifaceted CMR tasks with fewer annotations exemplifies its potential to streamline clinical workflows, reducing the need for extensive manual data labeling. It provides a scalable solution that democratizes access to high-performance models, promoting reproducibility, and accelerating clinical translation. By doing so, CineMA can help institutions overcome barriers related to data and resource limitations.

The comprehensive validation across diverse CMR tasks advocates the utilization of foundation models in other modalities or subspecialties of radiology, paving the way for prospecting new AI-driven enhancements in clinical monitoring and management. Subsequently, future research could extend CineMA's functionality to predict long-term patient outcomes or further optimize its computational efficiency without compromising accuracy.

Overall, CineMA's release as an open-source solution can invigorate collaborative advancements in cardiovascular AI research and practice, fostering a community-oriented progression in medical imaging innovation.