Embedded Files
Research Summary
Research Summary
My doctoral research focuses on developing advanced computational methods for cardiac magnetic resonance imaging, with a particular emphasis on left atrial analysis in patients with atrial fibrillation. I work on AI-supported methods for 3D LGE-MRI segmentation, thin left atrial wall analysis, conditional medical image synthesis, few-shot adaptation, and shape complexity measurement.
The long-term goal of my research is to support more reliable image-based assessment of atrial remodeling and improve treatment planning for atrial fibrillation.
Atrial Fibrillation and Atrial Remodeling
Atrial Fibrillation and Atrial Remodeling
Atrial fibrillation is a common sustained cardiac arrhythmia in which the atria lose coordinated electrical and mechanical activity. In a normal heart, electrical signals follow an organized pathway that supports coordinated atrial contraction. In atrial fibrillation, electrical activity becomes irregular and disorganized, often involving abnormal impulses around the left atrium and pulmonary veins.
Over time, physiological and pathological stressors such as atrial stretch, inflammation, and oxidative stress can contribute to fibrosis and structural remodeling of the atrial wall. This remodeling may alter tissue architecture and electrical conduction, creating a substrate that can sustain and worsen atrial fibrillation. In this way, atrial fibrillation and atrial remodeling can form a reinforcing cycle: abnormal rhythm promotes remodeling, and remodeled tissue increases the likelihood of persistent or recurrent arrhythmia.
Treatment Planning and Post-Ablation Assessment
Treatment Planning and Post-Ablation Assessment
Catheter ablation is a major treatment strategy for atrial fibrillation. The procedure aims to electrically isolate or modify arrhythmogenic regions, commonly near the pulmonary veins, by creating therapeutic scar tissue that blocks abnormal conduction pathways.
However, atrial fibrillation can recur when the underlying atrial substrate remains diseased, when fibrosis is extensive, or when abnormal conduction pathways are not fully eliminated. This makes the structural condition of the left atrial wall clinically important. A patient’s atrial substrate may influence treatment response, recurrence risk, and long-term rhythm outcomes.
LGE-MRI as a Window into the Left Atrial Wall
LGE-MRI as a Window into the Left Atrial Wall
Late gadolinium-enhanced magnetic resonance imaging provides a non-invasive way to visualize atrial tissue characteristics. In atrial fibrillation research, LGE-MRI is used to study left atrial anatomy, fibrosis, ablation-related scar, and post-treatment tissue changes. Hyperenhanced regions may correspond to fibrotic or scarred tissue, while non-enhanced regions may represent healthier myocardium.
This makes LGE-MRI valuable for studying atrial remodeling, but its full potential depends on accurate computational analysis. Before fibrosis burden, scar distribution, wall geometry, or shape-complexity markers can be measured, the left atrial cavity and wall must be identified reliably.
Imaging challenge: LGE-MRI provides clinically meaningful information, but quantitative analysis requires accurate segmentation of thin and complex atrial structures.
Research Pipeline
Research Pipeline
My work is organized as a sequential research pipeline. Each stage addresses a technical bottleneck that enables the next level of analysis. Accurate segmentation enables reliable shape measurement. Reliable shape measurement enables remodeling quantification. Remodeling quantification enables outcome modeling and risk stratification.
Stage 1 — Segmentation
Robustly segment the left atrial cavity and thin atrial wall from 3D LGE-MRI.
Stage 2 — Shape-Complexity Analysis
Extract quantitative markers of atrial remodeling, including geometric, radiomic, and fractal-based descriptors of wall structure.
Stage 3 — Outcome Prediction
Use imaging-derived biomarkers to study recurrence risk, treatment response, and clinically meaningful atrial fibrillation outcomes.
The logic is simple: if we can segment the left atrial wall reliably, we can quantify atrial remodeling more accurately; if we can quantify remodeling accurately, we can investigate which structural markers are associated with treatment outcomes; and if these markers are predictive, they may support more personalized atrial fibrillation care.
The Segmentation Challenge
The Segmentation Challenge
Quantitative analysis of atrial remodeling begins with segmentation. To measure wall shape, fibrosis distribution, scar burden, or surface complexity, the left atrial cavity and wall must first be delineated from the 3D LGE-MRI scan. This is technically challenging because the left atrial wall is thin, has limited contrast, and lies near surrounding anatomical structures with similar image intensity.
For deep learning models, this creates several difficulties: the wall occupies a small fraction of the image volume, boundaries are often blurred, pulmonary veins and the mitral valve introduce complex topology, and publicly available annotated datasets remain limited. These challenges make left atrial wall segmentation much harder than standard organ segmentation tasks.
Research Themes
Research Themes
Theme 1 — Cavity-to-Wall Transfer Learning for Thin-Wall Segmentation
Theme 1 — Cavity-to-Wall Transfer Learning for Thin-Wall Segmentation
Polished copy
Polished copy
A major focus of my research is improving left atrial wall segmentation by leveraging the more stable and easier cavity segmentation task. Because the left atrial cavity provides strong anatomical context for locating the surrounding wall, my work investigates cavity-to-wall transfer learning as a way to guide thin-wall segmentation in 3D LGE-MRI.
Project card
Project card
C2W-Tune: Cavity-to-Wall Transfer Learning for Thin Atrial Wall Segmentation
A transfer learning framework that uses left atrial cavity segmentation as an anatomical prior to improve thin atrial wall segmentation from 3D LGE-MRI.
What this contributes
What this contributes
Uses anatomical structure to guide wall segmentation.
Addresses the difficulty of learning thin-wall boundaries from limited data.
Supports the first stage of the broader remodeling-analysis pipeline.
Theme 2 — Few-Shot Adaptation for Limited Cardiac MRI Annotations
Theme 2 — Few-Shot Adaptation for Limited Cardiac MRI Annotations
Polished copy
Polished copy
Detailed wall annotations are difficult and time-consuming to obtain, and cardiac MRI data can vary across scanners, sites, and acquisition protocols. To address this, I study few-shot and meta-learning methods that enable segmentation models to adapt using only a small number of labeled 3D volumes.
Project card
Project card
Few-Shot Left Atrial Wall Segmentation via Meta-Learning
A gradient-based meta-learning framework designed to learn a transferable initialization for rapid adaptation to thin-wall segmentation with limited annotated data.
What this contributes
What this contributes
Reduces dependence on large labeled datasets.
Supports adaptation to new cohorts or acquisition conditions.
Makes left atrial wall segmentation more practical for real-world cardiac MRI research.
Theme 3 — Conditional 3D LGE-MRI Synthesis for Data Augmentation
Theme 3 — Conditional 3D LGE-MRI Synthesis for Data Augmentation
Polished copy
Polished copy
Another direction of my work investigates conditional generative models for 3D cardiac MRI synthesis. These methods generate realistic LGE-MRI volumes from anatomical or semantic conditioning maps, with the goal of increasing training diversity and improving downstream segmentation performance.
Project card
Project card
3D Conditional Image Synthesis of Left Atrial LGE-MRI
A comparative study of GAN- and diffusion-based models for generating synthetic cardiac MRI volumes from composite semantic masks.
What this contributes
What this contributes
Supports data augmentation when annotated datasets are limited.
Studies how synthetic images can improve segmentation robustness.
Connects image synthesis directly to downstream cardiac MRI analysis rather than treating generation as an isolated task.
Theme 4 — Shape Complexity and Quantitative Atrial Remodeling
Theme 4 — Shape Complexity and Quantitative Atrial Remodeling
Polished copy
Polished copy
Once the left atrial wall is segmented, the next step is to quantify its structural remodeling. I am interested in shape-complexity analysis, fractal-based descriptors, radiomic features, and multiscale geometric markers that may capture irregularity in atrial wall structure. These markers can provide a bridge between image segmentation and clinically meaningful outcome analysis.
Project card
Project card
Shape Complexity Analysis of the Left Atrium
Development of quantitative imaging biomarkers for characterizing atrial remodeling from segmented left atrial structures.
What this contributes
What this contributes
Moves beyond segmentation accuracy toward clinical interpretation.
Links wall geometry and remodeling to measurable imaging biomarkers.
Provides the foundation for recurrence-risk and outcome modeling.
Theme 5 — Geometry- and Topology-Aware Wall Modeling
Theme 5 — Geometry- and Topology-Aware Wall Modeling
Polished copy
Polished copy
Because the left atrial wall is a thin anatomical structure, segmentation methods should respect its geometry. I am interested in models that incorporate surface relationships, signed-distance representations, thickness-aware constraints, and topology-aware learning to better preserve wall anatomy.
Project card
Project card
Geometry-Aware Left Atrial Wall Segmentation
Exploring anatomically constrained segmentation methods that represent the atrial wall as a thin structure bounded by related inner and outer surfaces.
What this contributes
What this contributes
Encourages anatomically plausible wall segmentation.
Reduces unrealistic discontinuities or thickness errors.
Strengthens the reliability of downstream shape-complexity measurements.
Selected Publications
Selected Publications
Accepted Conference Papers:
Accepted Conference Papers:
C2W-Tune: Cavity-to-Wall Transfer Learning for Thin Atrial Wall Segmentation in 3D LGE-MRI
Accepted at AIME 2026. Preprint
Few-Shot Left Atrial Wall Segmentation in 3D LGE-MRI via Meta-Learning
Accepted at IEEE EMBC 2026. Preprint
Published Conference Papers:
Published Conference Papers:
3D Conditional Image Synthesis of Left Atrial LGE MRI from Composite Semantic Masks
Published in 2025 IEEE International Conference on Imaging Systems and Techniques (IST). Paper. Preprint
Research Impact
Research Impact
The broader aim of my research is to develop reliable and scalable cardiac MRI analysis methods that can support atrial fibrillation research and treatment planning. By improving left atrial wall segmentation, enabling adaptation with limited annotations, and extracting quantitative markers of atrial remodeling, this work contributes toward imaging-based tools for recurrence-risk analysis, treatment-response assessment, and personalized cardiac care.
Collaboration Interests
Collaboration Interests
I am interested in collaborations related to cardiac MRI, atrial fibrillation imaging, left atrial wall segmentation, medical image synthesis, few-shot learning, shape analysis, and clinically interpretable imaging biomarkers. I welcome discussions with researchers, clinicians, engineers, and industry partners working at the intersection of cardiac imaging and computational healthcare.
Page updated
Google Sites
Report abuse