First-hand information for everyone
Combining the two through agentic systems is a promising way to achieve comprehensive radiology reporting, yet existing systems are usually restricted to a single modality or body part.
A newly proposed radiology agentic framework standardizes how an orchestrator calls modular tools. The framework has three layers: resources, 22 tool functions organized into seven categories, and an orchestrator that plans, calls tools, and writes the report. When a tool is unavailable, the orchestrator can skip or flag that step.
To isolate tool availability and orchestration from tool error, the authors simulated tool outputs from a reference report. They evaluated the system on 1,059 closed‑source PARROT reports, a large multi‑modality report corpus, with the orchestrator receiving either a catalog based on FDA‑cleared radiology AI devices (1,989 classes) or an unlimited catalog that allows any class. The planner was given either (i) only the clinical indication (realistic setting) or (ii) the target report as an oracle.
Results, expressed as actionable recall in the RadMatch metric, show that with the cleared catalog and realistic planner the system recovers 0.290 findings. Providing the full target report boosts recall to 0.426. With an unlimited catalog the recall rises to 0.437 with realistic planning and 0.740 when the oracle report is used. The oracle gives a residual upper bound of 0.740, indicating a ceiling driven by tasks not expressible as tool calls.
Adding frequently missing tools also improves recall. Augmenting the catalog with 24 tools that were missing in at least ten reports increases recall to 0.434, closing 37 % of the gap to the unlimited catalog. Expanding to 99 missing tools boosts recall to 0.519 (74 % of the gap), and including 402 tools raises recall to 0.534 (80 % of the gap), although most additional calls yield few new positive findings.
An analysis of missed findings reveals that 61.8 % are catalog gaps—classes not present for the modality in the FDA‑cleared catalog. An additional 31.9 % are orchestration gaps—classes offered but not called or class‑free tools that could have surfaced them. Only 6.1 % require a new kind of tool and 0.2 % cannot be found by any imaging tool.
PARROT reports are fictional, and tool calls are simulated from the reference, so the results do not capture real‑world tool error or clinical variability. The study demonstrates that agentic systems can already compose today’s cleared radiology tools into a report, but coverage and orchestration remain key barriers to full recall.