From Research to Hospitals

This project had two halves. First I invented an algorithm, then I built the platform that put the lab's research in the hands of clinicians.

Part 1: The algorithm

The problem

Co-registration aligns the frames of coronary imaging video, so clinicians can compare what they see across images of the same artery. The lab's existing approach trained a model to predict the transformation parameters and the location of each frame. It had three problems:

• It overfit. Clinical data is scarce, and even with augmentation the model memorized the training set.
• It needed a GPU. The algorithm has to run on the medical device that records the data, and that device doesn't have one.
• It was slow. Co-registration took far too long to fit into the clinical workflow.

My decision

I replaced the learned model with a landmark-based algorithm. It defines landmarks in each frame, then aligns the frames by spline interpolation between the landmarks.

• No training, so no overfitting. The method doesn't depend on how much clinical data exists.
• CPU only. It runs on the imaging machine itself, inside the clinical workflow.
• Cross-platform. I made it run on macOS, Windows and Linux, because medical hardware runs all three.

I also built a frame-cleaning pipeline: CLIPSeg and SAM segment the medical equipment in each frame, and Stable Diffusion inpainting removes it.

Result

Co-registration became 40× faster, on CPU, on the device that records the data.

Part 2: The platform

The problem

The lab had several research prototypes, each useful on its own, but researchers and clinicians had no single place to use them.

What I built

• One full-stack platform (React, Node.js) that took 4 research prototypes to production
• The PyTorch models served through containerized APIs (Docker, Linux)
• An interface designed from clinicians' user stories, so it follows how they actually work

Result

The platform is used by researchers and clinicians at hospitals across the Boston area. I also onboarded and mentored 2 new engineers on the team.

What I learned

• A model isn't always the answer. With scarce data and no GPU, a classical method beat deep learning on every constraint that mattered.
• The deployment target shapes the design. Knowing the algorithm had to run on the imaging device decided the approach before any code was written.
• Research only matters once people can use it. The platform is what turned the lab's work into something clinicians use every day.