A multi-site diagnostics provider operating across urban and semi-urban locations, handling thousands of radiology scans monthly. The client operated a legacy reporting workflow where radiologists manually drafted reports for each scan — a process that was time-intensive, inconsistent in structure, and difficult to scale with growing scan volumes.
Report preparation consumed the majority of each radiologist's time, with a single diagnostic report taking 15–25 minutes to draft manually. Across 300+ daily scans, this created a systematic bottleneck that delayed patient results and reduced overall clinical capacity.
Rather than building a fully autonomous system, we designed an AI-assisted workflow that retained radiologist oversight while removing the repetitive drafting burden. The core principle: AI generates structured, editable draft reports; radiologists review, correct, and sign off.
This hybrid approach maximised adoption, maintained clinical accountability, and allowed the system to improve over time through radiologist corrections fed back into the model as training data.
Mapped existing reporting workflow across 3 sites. Identified 12 distinct report templates in use. Established baseline metrics for report time, error rate, and radiologist capacity.
Built a structured extraction pipeline to parse scan metadata, machine-generated findings, and radiologist shorthand into a standardised data schema.
Integrated GPT-4 API with domain-specific system prompts, constrained to approved clinical terminology. Output structured report drafts in the client's standard template format.
Built a React-based review interface where radiologists receive AI-drafted reports, make inline edits, and sign off — replacing blank-page drafting with structured review.
Corrections made by radiologists were logged and used to refine prompt templates and re-evaluate output quality bi-weekly, improving accuracy continuously post-launch.
80%
Reduction in report preparation time
3×
Increase in radiologist throughput
95%+
Report accuracy validated by clinicians
< 2 min
Average time from scan to structured report