HIPAA-Compliant AI Diagnostics
Building a secure, explainable AI radiology platform that flags anomalies 40% faster while meeting every HIPAA safeguard.
The Problem
A network of 14 radiology centres across the UK was facing a clinical backlog crisis. Radiologists were reviewing an average of 250 scans per day — well above the safe recommended volume — leading to increased error rates and delays in critical diagnoses. With NHS waiting list pressures growing, the business case for AI-assisted triage was clear.
The technical challenge was significant: any AI system handling patient-identifiable health data must satisfy strict HIPAA and UK GDPR requirements around data storage, access controls, audit logging, and model explainability. Off-the-shelf clinical AI solutions were either black boxes with no audit trail or so expensive they were inaccessible at this scale.
The Architecture
We built the platform on Azure's HIPAA-eligible services, with all patient data stored in Azure Blob Storage with customer-managed encryption keys (BYOK), scoped behind Azure Private Endpoints with no public internet exposure. All data in transit uses TLS 1.3; at-rest encryption uses AES-256. Full audit logs — who accessed what scan, when, and what the model returned — flow into Azure Monitor with 7-year immutable retention for regulatory compliance.
The AI model is a fine-tuned EfficientNet-B7 convolutional neural network, trained on a 225,000-image dataset of anonymised chest X-rays and CT scans, with transfer learning from NIH CheXNet weights. We used MONAI (Medical Open Network for AI) as the training framework, integrating DICOM-native preprocessing pipelines to preserve clinically relevant metadata.
Model outputs include both a classification (normal / flagged / urgent) and a Grad-CAM saliency map overlaid on the scan — giving radiologists a visual explanation of exactly which regions triggered the AI's assessment. This explainability layer was a non-negotiable requirement for clinical governance approval.
Technology Stack
The Outcome
Want similar results?
Let's discuss your project — we'll build the right solution together.