An AI Radiology Platform Re-Architected for Enterprise Hospital Networks

We began with a deep code audit against FDA 21 CFR Part 820 (Quality System Regulation) and the 2024 FDA Cybersecurity Guidance. The audit identified critical vulnerabilities in access management and data flow between services. Based on findings, we designed a security strategy with mutual TLS between microservices, network microsegmentation, and strict role-based access to PHI.

The most complex engineering challenge was the AI Bridge — a module that allows the vendor to hot-swap AI algorithms from third-party partners without platform downtime. We decomposed the monolith into microservices, each with its own deployment lifecycle. For critical pixel-processing components, we used Rust to guarantee memory safety and high throughput under concurrent load, which was essential for real-time DICOM manipulation where a buffer overflow could corrupt diagnostic data.

Instead of forcing radiologists to download entire imaging studies, we built a server-side rendering pipeline that delivers only the pixels currently visible in the viewer. The system pre-caches adjacent slices based on scroll direction prediction. This approach eliminated the latency bottleneck that had made the platform unusable for enterprise radiology workflows.

We designed logical tenant isolation using separate database schemas and encrypted S3 buckets per tenant, all running on shared computer infrastructure. Auto-scaling rules adjust capacity based on study volume. This architecture reduced the client’s per-hospital infrastructure cost by 30% compared to the previous single-tenant model while maintaining the data isolation required for FDA compliance.

Connecting to a hospital PACS previously required custom integration work for each client. We built a standardized gateway that speaks DICOMweb for imaging data and FHIR for clinical context. The gateway handles authentication, data mapping, and protocol translation automatically. Onboarding a new hospital dropped from eight weeks of engineering work to a one-week configuration process.

Substantial changes to a SaMD architecture can trigger the need for a new or updated 510(k) submission. We prepared the complete technical documentation package: Software Design Specifications, Risk Management File per ISO 14971, traceability matrices, and cybersecurity documentation aligned with the 2024 FDA guidance. This package formed the backbone of the client’s successful regulatory submission.