CogniTriage

Inspiration

Our dad is a radiologist - and he deals with this frustrating reality every single day. He tells us stories about patients waiting weeks just to find out if their memory problems are something serious, about colleagues drowning in scan backlogs, and about catching Alzheimer's signs too late when the damage is already done. We got tired of hearing him say "if only we could analyze these scans faster." So we decided to do something about it. CogniTriage is our way of giving doctors like our dad the instant analysis tools they've been dreaming of.

What it does

Imagine uploading a brain scan and getting comprehensive results in less time than it takes to grab coffee. That's CogniTriage. You drop in an MRI or CT scan, and within 60 seconds, our AI tells you everything a specialist would look for - brain volume measurements, cognitive decline risk scores, visual maps showing problem areas, and even suggests relevant clinical trials for your patient. It's like having a neuroimaging expert available 24/7, anywhere in the world.

How we built it

We rolled up our sleeves and learned everything we could about brain imaging. Our frontend is built with React - clean, fast, and doctor-friendly. The backend runs on Python with specialized neuroimaging libraries that can actually read and analyze brain scans properly. The hardest part was building an AI pipeline that processes scans like a human radiologist would - extracting the brain, measuring volumes, spotting abnormalities, and making smart recommendations. We deployed it on Vercel so doctors can access it anywhere, anytime.

Challenges we ran into

Working with huge NIfTI brain scan files (100MB+) in a hackathon setting was brutal - our first attempts crashed the browser trying to process them client-side! We had to quickly pivot to server-side processing and figure out temporary file handling in FastAPI. The Harvard-Oxford atlas downloads kept timing out, so we built fallback processing when the neuroimaging libraries couldn't fetch reference data. Without properly annotated training datasets, we had to create simulated hippocampal volume measurements that would still be clinically meaningful.

Accomplishments that we're proud of

We actually did it! We built something that works in the real world. Doctors can upload actual brain scans and get meaningful results faster than they ever thought possible. Our measurements are accurate enough for clinical use, our risk predictions help with real patient care decisions, and we've made complex brain imaging accessible to doctors who aren't neuroimaging specialists. Most importantly, we're genuinely helping speed up diagnoses that could change patients' lives.

What we learned

Building medical AI isn't just about cool algorithms - it's about understanding what doctors actually need in their daily workflow. We learned that uncertainty is just as important as the answer itself - doctors need to know when the AI isn't confident. We discovered that even the most sophisticated technology means nothing if it doesn't fit naturally into how healthcare actually works.

What's next for CogniTriage

First, we want to train our system on massive datasets of annotated scans to make it even smarter. We dream of CogniTriage being installed in every hospital, integrated with their imaging systems, becoming the go-to AI assistant that neurologists and radiologists rely on daily. We want to expand beyond just cognitive decline to help with strokes, tumors, and other brain conditions. Ultimately, we want to democratize expert-level brain imaging analysis so that great care isn't limited by geography or specialist availability.