Radiology Report QA and Peer Review Software: What to Look For (2026)

Disclosure and scope

Disclosure: RadQA is made by Zapx Labs, which publishes this blog. RadQA is at pilot stage, is demonstrated on a built-in demo dataset, and is a quality-assurance assistant, not a diagnostic or medical device. It has no published price; pricing is by conversation.

The guidance below comes from the Royal College of Radiologists (RCR), the Academy of Medical Royal Colleges (AoMRC) and the American College of Radiology (ACR). We read each document on 8 November 2026 and link them at the end. Guidance changes, so check the current version before relying on it.

Why sampling by hand misses things

In most departments, radiology QA is retrospective sampling done by hand. Someone pulls a set of reports, reads them against the images or the request, and records what they find. It works for learning, but it cannot see everything: a laterality conflict in an unsampled report, a critical finding that was never communicated, or a CT pulmonary angiogram acquired without the expected contrast can all pass unnoticed because nobody had time to look.

The standards ask for more than a sample. The RCR's Standards for interpretation and reporting of imaging investigations (third edition, October 2025) say providers should create standard operating procedures for turnaround times, communication of unexpected or time-critical findings, communication of addenda and report acknowledgement. The AoMRC's 2022 recommendations, Alerts and notification of imaging reports, say employing organisations are responsible for making sure fail-safe systems are in place and for auditing them regularly.

What the guidance expects you to check

These are the points most often turned into audit criteria. Each comes from the source document named.

• Alert categories (AoMRC, 2022): new cancer diagnoses or recurrences suspected from imaging, critical findings where management is time-critical, and significant addenda that may change management.
• Escalation (AoMRC, 2022): an alert that has not been reviewed and acted on should be escalated; a 48-hour interval before escalation is described as reasonable, varying with clinical urgency.
• Learning meetings (RCR, 2020): each trust should hold at least six radiology events and learning meetings (REALMs) a year, and every radiologist should attend at least half and contribute at least one case a year.
• No scoring in REALMs (RCR, 2020): scoring or categorising suspected errors in the meeting is not recommended, because REALMs are for learning rather than individual performance assessment.
• Peer review scales (ACR): RADPEER scores 1 (concur), 2 (an understandable miss) and 3 (a discrepancy that should be made most of the time), with a or b for unlikely or likely clinical significance; ACR says 2b, 3a and 3b should first go to internal review. ACR's page also announces RADPEER QI as its replacement programme.

How to catch laterality errors and other report defects

Report QA is the part of radiology QA most suited to automation, because the evidence is text. A laterality error is a report where the side is not consistent, for example a left-sided finding summarised as right-sided in the impression. Read by hand, it takes careful attention per report; written as a rule, the same check can run on every report as it arrives.

The useful pattern is a short list of explicit checks, each producing a flag a person can confirm in seconds because it quotes the sentence that triggered it.

• Laterality: the side named in the findings and the impression agree.
• Consistency: the impression does not contradict or leave out a significant finding.
• Critical-result communication: a critical or alert-category finding has a documented communication.
• Follow-up: findings that usually warrant follow-up, such as an incidental nodule, carry a recommendation.
• Template leftovers: no unfilled placeholders or default normal text that conflicts with the findings.

What to look for in radiology QA software

Whichever product you look at, these questions separate a QA tool you can govern from one you cannot.

• Coverage: does it check every study, or only a sample you choose?
• Breadth: report, protocol and image checks, plus peer review, or only one of these?
• Evidence: does every flag quote the sentence, or mark the region on the image, that caused it?
• Standard interfaces: DICOMweb (QIDO-RS and WADO-RS) for PACS, HL7 v2 over MLLP for RIS, and FHIR R4, rather than a bespoke export.
• Approval before write-back: can anything change RIS or PACS without a named person approving it?
• Data flow: where does report text or image data go when a model-backed check runs, and where are keys kept?
• Reproducible peer review: is case selection deterministic, so the sample can be explained and repeated?
• Analytics: QA score, turnaround, SLA breaches and the critical-communication gap, by modality rather than by individual by default.
• Positioning: does the supplier say plainly that it is not a diagnostic device?

Where RadQA fits

RadQA is a radiology quality-assurance workstation. It connects to a DICOMweb PACS, an HL7 v2 RIS feed over MLLP, a FHIR R4 server or a local DICOM folder, or starts from its built-in demo dataset. Studies and reports are checked as they arrive or on demand, combining deterministic rules with model-backed checks.

Four kinds of QA run on each study. Report QA covers laterality conflicts, impression against findings, critical-result communication, missing follow-up and template leftovers. Protocol QA covers missing series, contrast and slice-thickness mismatches, and body part and modality checks. Image QA covers markers, exposure, collimation, clipped anatomy, rotation and burned-in identifiers. Peer review is RADPEER-style, with deterministic sampling; RadQA is not affiliated with ACR's programme.

Findings land in a worklist. Every report finding quotes the sentence it came from, and image findings are drawn as regions on the pixels in the built-in DICOM viewer. A Claude-powered agent beside each case can explain the likely cause and propose next steps, such as drafting an addendum, and any write-back needs explicit approval. Analytics show volume, QA score, turnaround, SLA breaches and the critical-communication gap, filterable by modality and connector.

RadQA runs as a desktop app with local storage, and API keys stay in the operating system's keychain. Model-backed checks send the text or image needed for that check to the configured model provider, so it is not an offline-only tool.

What RadQA is not

• Not a diagnostic device: it flags likely issues for a person to review and does not report studies.
• Not an alerting system: it can show that a critical finding has no recorded communication, but it does not send clinical alerts or replace your organisation's fail-safe system.
• Not a replacement for REALMs or the chair's judgement on what to discuss.
• Not yet in routine use: it is a pilot, shown on demo data, with no department running it in production.

How to evaluate it

The RadQA site has a four-minute demo on demo data showing report QA catching an uncommunicated critical finding, protocol and image QA, and the agent drafting an addendum that waits for approval. To see it against your own questions, book a demo; the next step for a department would be running it on synthetic data, then on its own systems under local governance.