Healthcare · Computer Vision
HIPAA-ready AI radiology triage
Explainable medical imaging AI that flags anomalies faster while satisfying clinical governance and data safeguards.
- Review time reduction
- 40%
- Critical FN improvement
- 18%
- Daily scan volume
- 8,000+
- Inference
- <3s / scan
Problem
A network of radiology centres was reviewing ~250 scans/radiologist/day — above safe volume — with growing backlog pressure.
Any system touching identifiable health data needed encryption, private networking, audit trails, and model explainability. Off-the-shelf tools were either opaque or commercially unreachable.
Approach
Azure HIPAA-eligible stack: Blob with customer-managed keys, Private Endpoints, TLS 1.3 in transit, AES-256 at rest, immutable audit retention.
Fine-tuned EfficientNet-B7 on a large anonymised X-ray/CT set using MONAI, with DICOM-native preprocessing.
Outputs included triage class plus Grad-CAM saliency overlays so clinicians could see why a case was flagged — required for governance sign-off.
Outcome
Radiologist review time dropped ~40% via smarter worklists. Critical false-negatives improved ~18% vs unaided review. The platform cleared technical safeguard review and now processes thousands of scans/day with sub-3s inference on GPU endpoints.
Stack
- PyTorch
- EfficientNet-B7
- MONAI
- Azure ML
- DICOM
- Grad-CAM
- FastAPI
- React