Six papers in leading radiology journals within two years

The full record

Results by study

1. Diagnostic Accuracy and Clinical Value of a Domain-specific Multimodal Generative AI Model for Chest Radiograph Report Generation

   01View the paper: Radiology · 2025 , DOI link doi.org/10.1148/radiol.241476 (opens in a new tab)

   70.5%

   AI preliminary reports accepted without change

   Study detail+

   70.5% of AI-written reports were used without a single edit, against 73.3% for radiologist reports and 29.6% for GPT-4V in the same evaluation.

   A retrospective evaluation on a defined test set.

   Study design

   • Retrospective
   • 42 hospitals
   • Training 8,838,719 pairs
   • Test 2,145 radiographs
   • 4 readers

   Comparators in the same study

   Radiologist reports

   73.3%

   GPT-4V reports

   29.6%

   Result in full

   70.5% of model-generated reports were accepted without change, compared with 73.3% of radiologist reports and 29.6% of GPT-4V reports. Reports from the model ranked highest in 60.0% of evaluations. The authors concluded that the model showed potential for preliminary interpretation support.

   Limitations

   A retrospective evaluation across a defined test set. This is evidence about the studied model and setting, not proof of performance at every institution.

   Paper title

   Diagnostic Accuracy and Clinical Value of a Domain-specific Multimodal Generative AI Model for Chest Radiograph Report Generation

2. Value of Using a Generative AI Model in Chest Radiography Reporting: A Reader Study

   02View the paper: Radiology · 2025 , DOI link doi.org/10.1148/radiol.241646 (opens in a new tab)

   19.8s

   Mean reading time with AI preliminary reports

   Without AI34.2s

   With AI19.8s
   Study detail+

   Reading with an AI report brought mean reading time down from 34.2 to 19.8 seconds, and agreement and report quality rose in the same study.

   A sequential retrospective design on a public dataset, with variability among individual readers.

   Study design

   • Retrospective sequential MRMC
   • Public dataset
   • 758 radiographs
   • 5 readers

   Comparators in the same study

   Same readers, without AI preliminary reports

   34.2s

   Result in full

   Mean reading time changed from 34.2 to 19.8 seconds. Agreement and quality scores improved statistically, while the paper noted variability among individual radiologists.

   Limitations

   A sequential retrospective design using a public dataset. The observed time and quality changes should not be generalised to every workflow.

   Paper title

   Value of Using a Generative AI Model in Chest Radiography Reporting: A Reader Study

3. Clinical Validation of a Generative Artificial Intelligence Model for Chest Radiograph Reporting: A Multicohort Study

   03View the paper: Radiology · 2025 , DOI link doi.org/10.1148/radiol.250568 (opens in a new tab)

   88.4%

   AI reports rated acceptable (standard criterion)

   Study detail+

   Across ICU, emergency, health-checkup and outpatient cohorts, 88.4% of AI reports were clinically acceptable, against 89.2% for radiologist reports, a difference that was not statistically significant.

   The standard-criterion difference was not statistically significant, and under the stricter no-revision criterion AI reports were lower.

   Study design

   • Retrospective multicohort
   • ICU, emergency, health-checkup, public outpatient
   • 1,539 radiographs
   • 7 thoracic radiologists

   Comparators in the same study

   Radiologist reports · standard criterion

   89.2%

   AI preliminary reports · no-revision criterion

   66.8%

   Radiologist reports · no-revision criterion

   75.7%

   Result in full

   Under the standard criterion, acceptability did not differ statistically between AI and radiologist reports (88.4% vs 89.2%). Under the stricter no-revision criterion, AI reports were lower (66.8% vs 75.7%). In the compared cohorts, AI reports had higher sensitivity and lower specificity for referable abnormalities.

   Limitations

   Many reports required minor revision, and the surveyed radiologists did not regard the generated reports as a replacement for radiologist-written reports.

   Paper title

   Clinical Validation of a Generative Artificial Intelligence Model for Chest Radiograph Reporting: A Multicohort Study

4. Multimodal Generative Artificial Intelligence Model for Creating Radiology Reports for Chest Radiographs in Patients Undergoing Tuberculosis Screening

   04View the paper: AJR · 2025 , DOI link doi.org/10.2214/AJR.25.33059 (opens in a new tab)

   95.2%

   Sensitivity to TB-related abnormalities in AI preliminary reports

   Reader accuracy with AI: 91.3% → 93.5%

   Without AI91.3%

   With AI93.5%
   Study detail+

   Of the two radiologists reading tuberculosis-screening radiographs, one improved significantly in accuracy when reading with an AI preliminary report.

   A retrospective study of 800 radiographs from two public tuberculosis-screening datasets.

   Study design

   • Retrospective
   • 2 public TB-screening datasets
   • 800 radiographs
   • 2 readers

   Comparators in the same study

   Standalone AI report accuracy and localisation

   Lower than both readers

   Result in full

   Standalone AI report accuracy and localisation were lower than the readers'. Accuracy improved significantly with AI for one of the two readers.

   Limitations

   A retrospective study in which two readers read 800 radiographs from two public tuberculosis-screening datasets. Standalone AI report accuracy and localisation were lower than both readers'.

   Paper title

   Multimodal Generative Artificial Intelligence Model for Creating Radiology Reports for Chest Radiographs in Patients Undergoing Tuberculosis Screening

5. Detection of Interstitial Lung Abnormalities on Chest Radiographs: Diagnostic Performance of Radiologists and Artificial Intelligence

   05View the paper: AJR · 2026 , DOI link doi.org/10.2214/AJR.26.34686 (opens in a new tab)

   AUC 0.92

   Fibrotic ILA detection (CT reference standard)

   Mean 0.86 for five thoracic radiologists

   Study detail+

   Interstitial lung abnormality is hard to catch on a chest radiograph even for radiologists. In this study AIRead-CXR performed comparably to them.

   ILA can be difficult to detect on Chest X-ray, even for radiologists.

   Study design

   • 1,168 individuals
   • 41 CT-confirmed ILA
   • CT reference standard
   • Radiologists vs AI

   Result in full

   For ILA, a finding that is difficult to detect on Chest X-ray, AIRead-CXR performed comparably to radiologists.

   Limitations

   This record is topic-specific research. It is neither broad validation of AIRead-CXR nor a screening indication.

   Paper title

   Detection of Interstitial Lung Abnormalities on Chest Radiographs: Diagnostic Performance of Radiologists and Artificial Intelligence

6. Comparative Evaluation of Generative AI Models for Chest Radiograph Report Generation in the Emergency Department

   06View the paper: European Radiology · 2026 , DOI link doi.org/10.1007/s00330-026-12648-8 (opens in a new tab)

   84.5%

   Clinical acceptability of AI preliminary reports (highest of 4 comparator VLMs)

   Study detail+

   Of five medical-imaging VLMs evaluated under the same conditions, AIRead had the highest clinical acceptability and the lowest clinically significant disagreement.

   A single-center retrospective study with image-only model inputs.

   Study design

   • Retrospective
   • Single tertiary-center ED
   • 478 patients
   • 3 blinded readers
   • Image-only inputs
   • Same-day CT reference

   Comparators in the same study

   Radiologist reports

   74.3%

   The other four evaluated VLMs

   41.1–71.4%

   Result in full

   Among the five evaluated VLMs, AIRead had the lowest clinically significant disagreement and the highest clinical acceptability. Its hallucination rate was not statistically different from radiologist reports.

   Limitations

   A single-center retrospective study with image-only model inputs. Patient outcomes were not assessed, and human–AI collaboration was not directly evaluated.

   Paper title

   Comparative Evaluation of Generative AI Models for Chest Radiograph Report Generation in the Emergency Department