MRI Tumor Viewer
Brain tumor segmentation running entirely in the browser.
ML Engineer Intern, Wake Forest CAIR
- 90%
- classification accuracy
- 20min → 5min
- clinical processing time
- 1,250+
- patient cases
- 10TB+
- imaging data
The problem
Reviewing scans is slow and the bottleneck is human attention, not imaging. Cutting time-to-first-read matters more than a marginal accuracy gain, because the queue is what delays patients.
Approach
Train the segmentation model offline on GCP, then ship inference to the browser so a scan can be reviewed without uploading patient data anywhere.
Architecture
- Training pipeline in PyTorch on GCP across 50K+ scans, provisioned with Terraform for reproducibility.
- Model exported and quantized for TensorFlow.js so inference runs client-side on the reviewer machine.
- Viewer renders slice-by-slice with the predicted segmentation as an overlay the clinician can toggle.
- Firebase handles auth and durable metadata; scan pixels never leave the client in the browser path.
Design decisions
In-browser inference over a server endpoint
Medical imaging carries real privacy constraints. Running the model client-side removes an entire class of data-handling risk and eliminates upload latency on large volumes.
Overlay the prediction rather than replace the scan
The model assists a clinician, it does not decide. A toggleable overlay keeps the original image authoritative and makes the model auditable at a glance.
What was hard
- Quantizing to a browser-deployable size without losing clinically meaningful accuracy.
- Class imbalance — tumor voxels are a small fraction of any volume, so raw accuracy is a misleading metric and evaluation had to weight recall.
What I took from it
- The throughput win came from workflow placement, not model quality. A slightly worse model that loads instantly beat a better one behind an upload step.
Stack
Other work
- Product
Atriveo
Job-search platform with a 5.0-rated Chrome extension and live customers.
- Applied Research
FOMC Intelligence
NLP pipeline turning Federal Reserve communications into market signals.
- AI Systems
Legal RAG
Retrieval-augmented generation over legal filings, with citations that hold up.