What problem does it solve?
For a routine radiograph, the radiologist reads the image and then dictates or types the report from scratch or from a predefined template. The authors of the Northwestern Medicine study describe an AI draft as a starting point, much as an attending radiologist verifies and edits a trainee report. Radiology Partners, the largest radiology practice in the United States, describes the pressure as rising imaging demand and a growing capacity gap, with worklists that keep getting longer.
Documentation is one of the few parts of the read that can be shortened without changing who decides. Northwestern's authors add that faster documentation may ease shortfalls in imaging access and reduce burnout. The risk sits in the same place: a draft that reads well can pass review with an error in it, so the workflow has to keep the radiologist's judgment in charge.
How does it work?
- Send the image and context. When a radiograph is acquired, an inference server receives the image together with clinical data from the EHR, such as the reason for examination, the procedure type, the comparison study and the name of the reading radiologist.
- Generate the draft. A multimodal model writes free text report language. Northwestern Medicine pairs a vision transformer image encoder with a text decoder of 125 million parameters and trained it on its own institutional dataset. Radiology Partners applies a multimodal foundation model from Cognita Imaging.
- Place the draft in the reporting software. The draft appears in the fields of the radiologist's normal reporting tool, so there is no second screen to open. Northwestern integrated it with Epic and PowerScribe.
- Review, edit and sign. The radiologist checks the draft against the image, edits it and signs. No report goes out before that step.
- Check for urgent findings. Because the draft describes severity and chronicity in text, the same output can be scanned for findings that need action. Northwestern screened 97,651 radiographs for unexpected pneumothorax this way and reported a sensitivity of 72.7% and a specificity of 99.9%.
- Log and sample. Every draft, edit and signature is logged, and a sample of final reports goes to peer review against a rating scale for clinical accuracy and text quality.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Emerging
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Hours saved | Not pooled | at least 63 hours | 1 | 1 organization |
| Productivity gain | Too few to pool | 15.5% | 1 | 1 organization |
Value drivers: Employee productivity, Speed and cycle time.
Indicative value
A health system with 12 hospitals reading about 700,000 plain radiographs a year
USD 83,333 to USD 1.8 million
Radiologist time freed, valued at the cost of that time per year
How this is calculated
Formula: radiographs * shareDrafted * secondsPerReport * gain / 3600 * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Plain radiographs reported per year radiographs, radiographs per year | 400,000 | 700,000 | Editorial assumption. The Northwestern Medicine study datasets were derived from 299,164 radiographs between November 2023 and April 2024, so the high value is a full year of a health system of that size. Replace with your own volume. |
| Share of radiographs where a draft is used shareDrafted, fraction of radiographs | 0.5 | 0.9 | Editorial assumption for the exam types and sites you enable first. Replace with the share your pilot reaches. |
| Documentation time per radiograph without a draft secondsPerReport, seconds | 150 | 190 | The Northwestern Medicine study measured a mean of 189.2 seconds for radiographs reported without the model. The low value assumes a faster department. Replace with your own timestamps. |
| Documentation time saved when a draft is used gain, fraction of documentation time | 0.05 | 0.155 | The high value is the 15.5% documentation efficiency gain reported by Northwestern Medicine in a nonrandomized study. The low value is an editorial haircut for a first deployment. |
| Fully loaded radiologist cost per hour costPerHour, USD per hour | 200 | 350 | Editorial assumption, replace with your own cost of radiologist time. |
What it leaves out: A value of time, not cash, and only if the freed minutes are used for more reads or less overtime. It leaves out the cost of the model, the integration and the peer review, any gain from catching urgent findings sooner, and the added reading effort when a draft is poor.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Radiology Partners
United States · Healthcare · 2025
Radiology Partners, the largest radiology practice in the United States, built Mosaic Drafting inside its Mosaic Clinical Technologies division. The tool applies a multimodal foundation model to pre draft x-ray reports (later also head CT) for a radiologist to review, edit and sign. The model comes from Cognita Imaging, which Radiology Partners acquired. In July 2025 the company said Mosaic Drafting was being implemented under Institutional Review Board protocols while it pursued FDA approval, so the record is a pilot. No outcome from live use is published on the sources cited here.
No outcome disclosed.
Northwestern Medicine
United States · Healthcare · 2024
Northwestern Medicine trained its own multimodal generative model on its institutional radiology data. For a plain radiograph the model writes a draft report from the image and clinical data such as the reason for examination, and the draft appears in the radiologist's reporting software, where the radiologist verifies and edits it. The model ran across the 12 hospital health system from November 2023 to April 2024. In a prospective cohort study, drafts were associated with a 15.5% documentation efficiency gain, and peer review of 800 studies found no difference in clinical accuracy or textual quality.
- Productivity gain: 15.5%, November 15, 2023 to April 24, 2024, radiograph reports
"We found a 15.5% documentation efficiency benefit with no decrease in clinical accuracy on peer review, representing a net time savings of over 63 documentation hours over the study period, or a reduction from roughly 79 to 67 radiologist shifts required to provide coverage."
Claimed by: organization - Hours saved: at least 63 hours, over the study period, November 15, 2023 to April 24, 2024
"We found a 15.5% documentation efficiency benefit with no decrease in clinical accuracy on peer review, representing a net time savings of over 63 documentation hours over the study period, or a reduction from roughly 79 to 67 radiologist shifts required to provide coverage."
Claimed by: organization
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A large archive of radiographs with signed reports and the clinical context entered with each order
- A peer review rating scale and a pool of radiologists who can score drafts and final reports
- Time stamps for documentation time with and without a draft
Systems to integrate
- Picture archiving and communication system (PACS) or the imaging archive
- Reporting software and radiology information system
- Electronic health record, for the reason for examination and prior studies
- Peer review or quality assurance tool
- Identity and access management for the inference service
Complexity: High
The model needs a paired archive of images and signed reports from your own department, or a foundation model that has been validated on your scanners and patient mix. The draft then has to appear inside the reporting software and the EHR, and the department needs a peer review process and a regulatory position before it scales.
- 1
Start with one exam type and one site
Pick a high volume exam such as a chest radiograph, where findings follow familiar patterns and a radiologist can judge a draft quickly. Leave cross sectional imaging for a later wave.
- 2
Build the dataset and the rating scale first
Pair images with signed reports, split them by patient so the same person never appears in training and evaluation, and agree on a rating scale for clinical accuracy and text quality before the first draft is written.
- 3
Put the draft where the radiologist already works
Insert it into the fields of the existing reporting tool as editable text. A separate window that has to be opened will be ignored on a busy shift.
- 4
Run a prospective comparison
Measure documentation time on studies with and without a draft in live reads, sample the final reports for peer review, and log every draft and edit. Northwestern's study used this design.
- 5
Settle the regulatory path before you scale
Ask the device regulator whether the drafting software counts as a medical device in your market. Radiology Partners said in July 2025 that its tool ran under Institutional Review Board protocols while it pursued FDA approval.
Guardrails
- The radiologist verifies, edits and signs every report; a draft is never released to a referring clinician on its own
- Drafts are enabled only for the exam types and sites that were validated
- Every draft, edit and signature is logged and traceable to the model version
- A fallback to the normal template when the model is down or returns nothing
KPIs to instrument
- Documentation time per study with a draft against without one, by radiologist and exam type
- Peer review score for clinical accuracy and text quality on final reports, with and without a draft
- Share of drafts accepted with little editing, and the types of error radiologists correct
- Sensitivity and specificity of any urgent finding flag against the signed report
Human in the loop
The radiologist owns the final interpretation and signs the report. A named clinical lead owns model performance, reviews the peer review sample every month and decides which exam types and sites are added.
Common failure modes
- Automation bias
- A fluent draft passes review with an error in it, because the radiologist reads the text before the image. Sample final reports for peer review and track the error types.
- Performance drift on new equipment or populations
- The model was trained on one department's scanners and patient mix and degrades elsewhere without any visible signal. Validate on local data and monitor accuracy by site.
- Drafts that are ignored
- Radiologists dictate over the draft because it needs too much editing, so no time is saved. Track acceptance and edit distance and retrain on the corrected reports.
What are the risks and rules?
EU AI Act
High risk
Software that creates diagnostic report text from a medical image is likely to be medical device software under the EU Medical Device Regulation. A product that needs third party conformity assessment under that regulation is high risk through Article 6(1) and Annex I of the EU AI Act, whether or not it appears in Annex III. That Annex I route applies from 2 August 2028 (Article 113(c)). Until then Article 4 (AI literacy) and Article 5 (prohibited practices) already bind the hospital as a deployer. The radiologist's signed interpretation stays a human decision.
Rules that apply
Guidance
- Regulation (EU) 2017/745 on medical devices (European Union, Europe). Software that provides information used for diagnostic decisions can be a medical device, and its risk class decides whether a notified body must assess it before CE marking.
- Article 6: Classification Rules for High-Risk AI Systems (European Union, Europe). A product that is itself a medical device requiring third party conformity assessment, or a safety component of one, is high risk under the EU AI Act.
- List of Artificial Intelligence-Enabled Medical Devices (US Food and Drug Administration, North America). FDA's list of AI enabled devices authorized for marketing in the United States. Check it, and the device's cleared indications, before relying on a drafting tool outside a research protocol.
Controls to put in place
- Documented intended use and the list of exam types and sites the model is validated for
- Radiologist signature on every report, with the draft and the edits kept in the audit trail
- Monthly peer review sample and performance report by site and exam type
- Model version control and a change process that includes revalidation on local data
Frequently asked questions
- How much time does AI report drafting save?
In a prospective cohort study of 11,980 model assisted radiograph interpretations at Northwestern Medicine, documentation efficiency rose by 15.5% and the authors calculated a net saving of over 63 documentation hours across the study period. The study was nonrandomized and covered one health system, so treat it as a first measurement rather than a benchmark.
- Does the AI replace the radiologist?
No. In both deployments on this page the radiologist reviews, edits and signs the report. Northwestern's peer review of 800 studies found no difference in clinical accuracy or text quality between model assisted and other reports.
- Should we train our own model or use a foundation model?
Both routes appear in the evidence. Northwestern Medicine trained a small model on its own institutional reports. Radiology Partners applies a multimodal foundation model that it acquired with Cognita Imaging. A model trained on your own data fits your report style, but it needs a large paired archive and a validation effort; a foundation model needs validation on your scanners and patients before use.
- Is a report drafting tool a medical device?
Often yes. Software that produces diagnostic text from an image is likely to count as medical device software in the EU and the United States, which is why Radiology Partners described its drafting tool as running under Institutional Review Board protocols while it pursued FDA approval. Ask your regulator before moving beyond a research protocol.
How to cite this page
Blits.ai AI Use Case Library, "AI drafting of radiology reports for radiologist review", last verified 14 October 2026, https://www.blits.ai/ai-use-cases/radiology-report-drafting-assistant. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 14 October 2026: First published