MedAI Diagnostics

Challenge

A hospital network group faced a growing backlog of medical imaging reviews caused by specialist availability constraints. Radiologists were reviewing thousands of scans per week under time pressure, and the group needed a way to surface high-priority cases without introducing data governance risk or compromising diagnostic reliability.

Approach

We worked closely with the clinical team to define the specific anomaly classes most important for triage and to understand the HIPAA and institutional governance constraints. We designed an on-premise inference architecture from the outset — no patient imaging data would leave each hospital's environment. Model training was conducted on a de-identified dataset approved by each institution's review board.

Solution

We developed a custom convolutional neural network trained on de-identified DICOM imaging data, with a Grad-CAM explainability layer that highlighted regions driving each classification — important for clinical trust. The model was packaged as a FastAPI service deployed within each hospital's infrastructure via Docker. A PACS integration layer routed flagged scans to a priority review queue in radiologists' existing workflow. Full audit logging was implemented to satisfy HIPAA requirements.

Results

• 94.3% anomaly detection accuracy on the held-out validation set
• Deployed across 12 hospital network nodes
• Zero HIPAA compliance issues identified during third-party audit
• Estimated 60% reduction in manual triage time for the radiology team