Area D—Applied Generative AI
Report Assistant — AI Radiology Reporting Suite
Flagship GenAI · eight sub-agents · six releases
An eight-agent LLM system radiologists use to draft reports — measured, not asserted.
- faster report turnaround
- 35%
- faster report turnaround
- false-positive recommendations
- 0
- false-positive recommendations
01/Problem
Writing the Impression is among the most demanding parts of reporting and where errors matter most. Automating it means beating hallucination, handling radically different rules per exam type, respecting each radiologist's personal style, and measuring output at scale before anyone trusts it.
02/Approach
Eight sub-agents shipped across six releases: Impression (with five body-region variants), Recommendation, Corrector, Prompt Insertion, Breast (BI-RADS-aligned), Report Preference, Thyroid (TI-RADS) and Lung (Lung-RADS). An orchestrator resolves the active preference set, routes to the right agent and runs the corrector in parallel — three agent calls per request, not a serial chain of eight. A three-tier preference hierarchy (reading physician › facility › facility group) lets radiologists keep their own voice. Drafting retrieves prior report context before generation. Human-in-the-loop by construction: the radiologist remains responsible for every signed report.
03/Outcome
- Recommendation agent: 94.8% accuracy (95% CI 92.4–97.0), 100% precision, zero false positives, hallucination 0.000 — it never invented an undictated recommendation.
- Impression agents: hallucination 0.005–0.025, correctness 0.72–0.84, similarity 0.59–0.75 across all variants.
- Structured reporting: 100% category accuracy — Lung-RADS on 560 reports, TI-RADS on 735 nodules, BI-RADS-aligned breast on 499 studies.
- Prompt directives 99% accurate (F1 0.96); Corrector 2.0 at 81% detection sensitivity vs 70% for the alternative model.
- Median latency under a second for impression and recommendation; API cost $0.015–$0.47 per 100 reports, tracked per release.
- 35% reduction in report turnaround time in live clinical workflow (internal business reporting).
04/My role
- Architected the eight-agent decomposition so each capability could be evaluated and iterated independently.
- Specified the orchestration — preference extraction, routing, parallel correction — holding each request to three agent calls.
- Designed the three-tier preference hierarchy that made the product adoptable.
- Required generative output be measured before release: a benchmark of 330 reports stratified by sex, age, facility group, state, modality and radiologist, scored by an LLM judge with bootstrapped CIs.
- Ran model selection as a measured procurement decision, weighing accuracy against latency and cost per report.