Integrating Deep Learning, Spatial Computing, and Physics Simulations for Digital Cardiology
Name
Kadry-kkadry-phd-hst-2026-thesis.pdf
Size
101.84 MB
Format
Adobe PDF
Checksum (MD5)
dc19649ed63c98023d2f1e394ecde5d1
Author(s)
Kadry, Karim
Advisor(s)
Edelman, Elazer R.
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
Biophysical modeling offers the potential for virtual interventions across a broad array of cardiovascular interventions and is increasingly important as device sophistication rises and inclusion criteria broaden. Virtual interventions have been increasingly integrated within digital cardiology workflows by solving partial differential equations on complex anatomical structures, enabling the extraction of clinically relevant treatment outcomes. In medical workflows, they can support the personalization of interventions to patient-specific anatomy; in engineering workflows, they can enable the evaluation of device designs across virtual cohorts; and in scientific workflows, they can reveal relationships between anatomic features and device outcomes. This thesis starts from the premise that virtual interventions, despite exhibiting clinical-grade accuracy, have yet to realize their full potential within digital workflows due to anatomical complexity. For example, simulation-ready anatomical models remain labor-intensive to reconstruct, limiting the scale of 3D datasets. Moreover, high-fidelity simulations remain too slow for rapid medical decision-making or device-design optimization. Lastly, anatomic variation between patients is complex, hampering the extraction of causal, rather than correlational, relationships between structure and device outcomes. I address these limitations by leveraging the latest advances in differentiable programming, spatial computing, deep learning, and physical simulation. First, I develop geometric processing and meshing tools to reconstruct simulation-ready virtual twins from multi-component coronary segmentations predicted through deep learning. Second, I develop a scalable virtual trial engine for patient-specific coronary angioplasty and train deep-learning models to rapidly emulate virtual interventions. Third, I develop anatomical diffusion models that are programmable with respect to geometry and topology, enabling counterfactual simulations that probe causal relationships between anatomical features and interventional outcomes. Together, these tools contribute towards digitally closing key feedback loops within cardiovascular medicine, engineering, and science.
MIT Department
Harvard-MIT Program in Health Sciences and Technology
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
