Abstract
Introduction
Heart failure (HF), a complex clinical syndrome characterized by the heart’s inability to pump blood efficiently, affects approximately 64 million people globally (Shahim et al., 2023). HF is associated with high mortality, poor quality of life, and remains a substantial burden on the healthcare system. There are many subtypes of HF, though the prevalence of heart failure with preserved ejection fraction (HFpEF) is rapidly increasing. The understanding of HFpEF is limited and there has been minimal success in therapy development, so there has been an increased focus on accelerating HFpEF research. Preclinical animal models are a vital tool in cardiac research as they provide a means to study disease mechanisms and test therapies in a living organism that has a structurally and functionally similar heart structure. This preclinical research often uses cardiac magnetic resonance imaging (MRI): an imaging modality that provides detailed cardiac structure and functional metrics. However, cardiac MRI analysis is time-intensive, error-prone, and inconsistent across researchers. With the increasing use of artificial intelligence (AI) to improve clinical efficiency, there is potential to streamline cardiac MRI analysis through automation techniques, though this development must be optimized for implementation across varying resource levels.
Technical Project
To analyze CMR images, researchers are currently required to manually trace the myocardium, a process called segmentation. This process is inefficient as it requires immense precision and is therefore time-consuming. Additionally, different researchers can segment the same cardiac MRI image differently, leading to inconsistencies in the cardiac metrics provided by the analysis. The goal of the technical component of my thesis was to design and implement AI-based deep learning models to automatically segment images from two cardiac MRI methods: Displacement Encoding with Stimulated Echoes (DENSE) and Whole-Heart Cine. By automating the segmentation process, this project aimed to enhance the precision and reproducibility of preclinical cardiac MRI data with an acceleration of the overall analysis pipeline.
The creation of the two deep learning models required a great amount of preprocessing, including filtering through cardiac MRI datasets to create diverse training datasets. The models were created using TensorFlow, a platform for machine learning model creation, in Python and were continuously trained on various hyperparameter combinations. Once our models reached an optimal performance, measured by a dice score ≥ than 0.85, we worked to create a pipeline to integrate these models into the current software for cardiac MRI analysis. Overall, this technical project accelerates preclinical cardiac MRI analysis, allowing for further research and therapy development for diseases like HFpEF.
STS Research Paper
While AI usage is healthcare is growing in higher-income countries, low- and middle-income countries (LMICs) lag in implementing these technologies, furthering healthcare access gaps. Although AI-based medical technologies have immense potential to improve healthcare systems in LMICs, their implementation has many underlying complexities. In the STS research component of my thesis, I performed a literature review to investigate the challenges and barriers LMICs face when it comes to adopting AI-based radiology tools, while proposing strategies for their adoption using a sociotechnical systems theory STS framework. Throughout my research, I examined case studies where AI clinical tools were tested in lower resource areas to understand the challenges of implementation while also considering the ways they can be adopted successfully.
Conclusion
The technical and STS components of my thesis are intertwined. When developing new medical technology, such as automated cardiac MRI segmentation models, engineers must consider how these technologies can be applied in areas different than those with unlimited resources. An ethical engineer considers the widespread use of the new technology and ensures existing inequities are not exacerbated. Without considering the environment and the users, known as the social system, a technical system cannot be efficient. My STS research taught me that the process for implementing AI-based radiology tools, such as the segmentation models designed in the technical project, will look different depending on the users, environment, and resources available – an ethical dimension that all engineers must consider.
Acknowledgements
Firstly, I would like to express my immense gratitude to my technical project team members: Esha Dua, Ethan Helicke, and Brian Warden. I’m also grateful to my technical advisors, Dr. Frederick Epstein and Thomas Skacel, for their guidance and support in my technical work. Finally, I'd like to thank my STS advisor, Dr. Richard Jacques, for his guidance in the compilation of this thesis.
