Research

MRI scans take forever. We all know this. So I built an unrolled reconstruction framework that combines good old model-based optimization with learned priors — basically giving the reconstruction a physics cheat sheet. The result? We got structural similarity scores around 0.96 even when throwing away most of the data (aggressive undersampling).

Here's the thing: models trained on one scanner often fall apart on another — different acquisition settings create invisible domain gaps. I noticed this and came up with a feature-distance-guided reweighting trick that helps the model generalize to unseen domains without ever needing labeled target data. It's like teaching someone to drive in any car, not just the one they practiced in.

On top of that, I trained a self-supervised representation learning model (DINO-style, if you're curious) that learns transferable features usable for both reconstruction and whatever downstream task you throw at it — segmentation, classification, you name it.

Geometric distortions in MRI are a pain — especially in prostate imaging, where even small warping can throw off a diagnosis. I thought: what if we use the distortion itself to guide the fix? That became the core idea behind DGR (Let Distortion Guide Restoration).

I built a physics-informed generative framework where a forward simulation shows the model what distortion looks like, and a diffusion-based model learns to reverse it. Then I paired this with a CNN-based geometric corrector, so the diffusion model handles the fine details while the CNN does the heavy structural lifting. They work together beautifully across a wide range of distortion magnitudes.

The numbers back it up: NMSE dropped by more than 0.1, and radiologists scored the corrected images about 1 point higher on average compared to existing methods. Not bad for making distortions work for us rather than against us.

Modern MRI pipelines involve a ridiculous number of steps: reconstruct, denoise, segment, quantify, generate a report. Doing all of this manually is slow, but AI agents kept failing when workflows got long. So I designed BCER — a modular agent that splits the brain work from the busy work.

The Brain figures out what to do. The Cerebellum figures out how — it resolves abstract references ("case.input", "node.id.field") into actual file paths at runtime, so the agent stops hallucinating broken dependencies. The Extremities execute each step in a sandboxed environment. And the Reflector catches failures locally instead of restarting the entire workflow from scratch.

The jump was dramatic: from 22% success rate with a standard ReAct agent to 93% on cardiac reporting tasks. I tested it across brain, prostate, and cardiac datasets — 8 different tasks, over 1,000 cases — and BCER hit a 99% end-to-end completion rate. Honestly, it felt like finally getting a Rube Goldberg machine to actually pour a cup of coffee.

This was one of my first tastes of building a complete hardware-to-cloud system. I designed a wearable health monitoring device that pulls together optical and electrical sensor data — heart rate, SpO₂, the works — and built the entire pipeline from the embedded firmware up through wireless streaming to cloud-based analysis.

It was equal parts circuit design, signal processing, and learning to debug firmware at 2am. But seeing real-time physiological data streaming from a tiny board to a dashboard felt like magic. This project really cemented my love for medical technology.