projects
Selected research contributions showing the technical arc of my work in computational medical imaging.
Projects that demonstrate the arc of my research program
These projects show how my work moves from clinically grounded problem formulation to technical innovation in inverse problems, generative modeling, and robust medical imaging.
Through-plane super-resolution for anisotropic MRI
A central theme of my research is how to recover missing spatial detail in clinical MRI without relying on unrealistic training assumptions. This line of work addresses a common clinical setting in which scans have high in-plane resolution but thick slices or slice gaps through plane.
Problem
Standard super-resolution methods often assume access to clean paired high-resolution targets or synthetic degradations that do not match how clinical MRI is actually acquired.
Contribution
I developed acquisition-aware, internally trained, and cycle-consistent approaches that explicitly model the MRI imaging operator, allowing reconstruction methods to learn from the scan itself and better reflect the physics of data acquisition.
Why it matters
This work shows how super-resolution can be made more realistic, more robust across settings, and less dependent on idealized supervised pipelines, helping lay the groundwork for clinically credible image enhancement.
Zero-shot and internally trained reconstruction
Across several projects, I study how imaging models can reduce dependence on large external paired datasets by learning directly from structure already present within an image or acquisition.
Problem
Supervised medical imaging pipelines often inherit biases from their training distributions and can fail when applied to new scanners, protocols, or clinical populations.
Contribution
I developed zero-shot and scan-specific learning methods that train on internal image structure rather than only on external datasets, enabling flexible reconstruction even when matched training data are unavailable.
Why it matters
This work advances a broader research agenda in which medical imaging systems are designed to adapt to local structure and acquisition context, rather than relying exclusively on fixed training corpora.
Generative priors for MRI inverse problems
I am interested in how modern generative models can be used in medical imaging without sacrificing reliability. This motivates my work on diffusion-based and generative formulations for inverse problems.
Problem
Generative models can capture rich image structure, but naively using them for reconstruction risks producing outputs that look plausible while departing from the actual acquired measurements.
Contribution
I study data-consistent generative inference, including diffusion-based priors that explicitly enforce agreement with measured observations while still leveraging learned image distributions.
Why it matters
This work addresses one of the central obstacles to deploying generative models in medicine: how to benefit from expressive priors without enabling clinically unsafe hallucinations.
Large-scale 3D brain MRI curation for generative modeling
Building useful generative models for medical imaging requires not only model innovation, but also careful dataset construction and evaluation. I treat this as a research problem in its own right.
Problem
Medical generative modeling is limited by fragmented datasets, variable preprocessing, and a lack of evaluation standards tied to anatomy and downstream utility.
Contribution
I have worked on curating and analyzing large-scale 3D brain MRI datasets for diffusion-model development, with emphasis on anatomical plausibility, data quality, and usefulness for reconstruction tasks.
Why it matters
This work strengthens the empirical foundation for medical generative imaging and supports a more rigorous pipeline from dataset design to clinically meaningful model evaluation.
Robustness across scanners, sites, and populations
Across reconstruction, harmonization, and downstream analysis, a recurring goal of my work is to reduce sensitivity to changes in scanner, site, acquisition protocol, and patient population.
Scientific question
How can we design imaging algorithms that remain reliable when the data distribution changes in ways that are inevitable in real-world clinical deployment?
Approach
I combine acquisition-aware modeling, robust learning, and structured priors to reduce unwanted sensitivity to scanner and protocol variation.
Program-level significance
This theme connects my current work to a broader faculty agenda centered on trustworthy and deployable computational imaging.
Cross-modal collaborations in imaging AI
While MRI is the main technical center of my work, my collaborations span CT, OCT, pathology, ultrasound, and multimodal imaging. These collaborations broaden the questions I ask and reinforce a larger view of computational imaging as a discipline.
Role of collaboration
Working across modalities helps identify which ideas are MRI-specific and which reflect deeper principles in inverse problems, representation learning, and robust image analysis.
Research value
These collaborations extend the reach of my methods while also informing future directions in multimodal and multi-image computational imaging.
Faculty perspective
They also illustrate the collaborative style of lab I aim to build: technically deep, clinically connected, and open to new imaging domains.