AI Radiology Copilot

AI Radiology Copilot covers one chest X-ray workflow from start to finish. It receives an imaging study, checks that the patient and order match, confirms that all required images are present, and displays the images alongside prior studies and clinical context. It then runs AI analysis, lets the radiologist review each flagged finding, drafts a report from the reviewed findings, and once an authorized radiologist signs, delivers the report and tracks that it was received.

It is built for the whole reading team. Radiologists work from a prioritized worklist and a full medical-image viewer with AI findings overlaid on the image. Trainees can draft preliminary reports but can't sign them. Technologists resolve identity mismatches and incomplete studies. Administrators manage users and access without seeing clinical data, and auditors can verify the integrity of the audit trail. It is designed for hospital radiology departments, imaging centers and teleradiology groups.

AI is treated strictly as decision support. Every finding the model flags shows its score, threshold and highlighted image region, and a radiologist must accept, edit or reject it. Reports are drafted only from reviewed findings, and only a credentialed radiologist with signing authority can sign. The system also publishes exactly which images it supports; anything outside that list, such as a pediatric study, a CT scan or a non-DICOM file, gets an explicit reason code instead of being analyzed silently. Critical results such as a collapsed lung (pneumothorax) are flagged, and their communication is documented with read-back. Reports are delivered in formats hospitals already use: FHIR DiagnosticReport and HL7 v2.

The security and data-integrity design follows clinical-system standards: per-site permissions, database row-level security, restricted-patient access grants, a hash-chained audit trail that makes tampering detectable, original images stored unchanged and identified by cryptographic hash, and durable background jobs with retries. The bundled AI model and images are synthetic, and the demo data deliberately includes hard cases: a critical finding, missing views, identity mismatches, unsupported studies, and a second site that the first site's users cannot see. Clinical use would require model validation, real system integrations and governance decisions, all documented in a deployment-inputs guide.