NovaMed — AI Diagnostic Assistant

Background

NovaMed’s research team had built a promising chest X-ray analysis model in PyTorch — 91% accuracy on their validation set. But a model in a Jupyter notebook is not a clinical tool.

They needed:

1. A production ML serving pipeline with <300ms inference
2. A radiologist-facing UI integrated with their PACS system
3. HIPAA compliance and audit trails
4. A feedback loop to keep the model improving

What we built

ML serving pipeline

We exported the PyTorch model to ONNX and served it with ONNX Runtime behind a FastAPI service. Inference time dropped from 1.4s (PyTorch) to 210ms (ONNX) with no accuracy loss.

The pipeline:

DICOM upload → Preprocessing → ONNX Runtime → Heatmap generation → Result store

We built a preprocessing layer that normalises DICOM images to the format the model expects, handling the various DICOM flavours from different scanners.

Radiologist UI

A clean React application showing the original X-ray alongside the model’s heatmap overlay, with a confidence score and a pre-populated finding report. Radiologists can accept, modify, or reject the AI findings — all interactions go back into the training dataset.

HIPAA compliance

• All data encrypted at rest (AES-256) and in transit (TLS 1.3)
• De-identification pipeline strips PHI before images leave the hospital network
• Full audit log of every access and every AI inference
• Business Associate Agreement with all infrastructure providers

Continuous learning

The feedback loop is the most valuable part: every radiologist correction is a labelled training example. The model retrains weekly and is evaluated against a held-out test set before deployment. Accuracy has improved from 91% to 94.2% over 6 months of operation.