Moderate evidence
More
AI second reader cut unnecessary biopsy decisions in a 12-doctor throat imaging study
Study title: AI-Assisted Ni Classification in Nasopharyngeal NBI: Multicenter Validation and Reader Study
- Evidence type
- Controlled or validated comparison
- Sample
- n = 5,404
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-3035cite or check it
Study title: AI-Assisted Ni Classification in Nasopharyngeal NBI: Multicenter Validation and Reader Study
Researchers built an AI system that sorts nasopharyngeal narrow-band images into five Ni classes and tested whether it helps doctors agree on diagnoses and avoid unneeded biopsies. In a reader study with 12 physicians and 200 images, AI assistance lowered potentially unnecessary biopsy decisions from 21.6% to 10.3% among junior physicians and from 13.7% to 8.0% among senior physicians.
Read the study abstract
To develop and validate an artificial intelligence (AI) system for standardized Ni classification of nasopharyngeal narrow-band imaging (NBI) and to evaluate whether AI assistance improves physician diagnostic consistency and biopsy-related decision making in a reader-study setting. In this multicenter diagnostic study, 5404 high-quality nasopharyngeal NBI images were retrospectively collected. Lesion segmentation and five-class Ni classification networks were trained independently. Performance was evaluated in…
The gain
In a randomized crossover study of 12 physicians reading 200 nasopharyngeal images, AI assistance cut potentially unnecessary biopsy decisions from 21.6% to 10.3% for junior physicians and from 13.7% to 8.0% for senior physicians.
In brief
Agreement between doctors also rose. Fleiss kappa went from 0.610 to 0.788 for junior physicians and from 0.744 to 0.870 for senior physicians, with both changes statistically significant. The authors say the results support further evaluation of the system as a second-reader tool.
The model itself was trained on 5404 retrospectively collected images and held up in outside testing, with accuracies of 0.889 and 0.873 in two independent external cohorts. The reader study used images from a single centre, CHSZCAMS, so real clinic performance remains to be shown.
Main points
- The study collected 5404 high-quality nasopharyngeal NBI images retrospectively across multiple centres.
- Separate networks were trained for lesion segmentation and five-class Ni classification.
- In two independent external cohorts, accuracy was 0.889 and 0.873, with AUCs of 0.976 and 0.967.
- In recorded-video validation, accuracy was 0.888 and AUC was 0.974.
- A prospective randomized multi-reader crossover study had 12 physicians read 200 images from CHSZCAMS.
- Agreement among junior physicians rose from a kappa of 0.610 to 0.788; among senior physicians from 0.744 to 0.870.
The full story
The team developed and validated an AI system for standardized Ni classification of nasopharyngeal narrow-band imaging. They trained lesion segmentation and five-class Ni classification networks independently on 5404 retrospectively collected images.
Performance was checked in an internal test set, two independent external cohorts and a recorded-video validation set. External accuracies were 0.889 and 0.873 with AUCs of 0.976 and 0.967, and video accuracy was 0.888 with an AUC of 0.974.
In the prospective randomized crossover reader study, 12 physicians interpreted 200 images with and without AI help. Junior and senior physicians both agreed with each other more, and both made fewer potentially unnecessary biopsy decisions.
The authors conclude that the tool showed consistent performance across validation cohorts and improved agreement across experience levels, supporting further evaluation as a second-reader tool.
Sources
1 on file- The Laryngoscopedoi:10.1002/lary.71067
Talk it over
Potentially unnecessary biopsy decisions fell from 21.6% to 10.3% for junior physicians and from 13.7% to 8.0% for senior physicians. Who benefits most from the AI second reader, and who might benefit least?
AI second reader cut unnecessary biopsy decisions in a 12-doctor throat imaging studyhttps://truvace.com/item/ai-assisted-ni-classification-in-nasopharyngeal-nbi-multicenter-validation-and-r
Potentially unnecessary biopsy decisions fell from 21.6% to 10.3% for junior physicians and from 13.7% to 8.0% for senior physicians. Who benefits most from the AI second reader, and who might benefit least?
Answer in the threadThe 12 physicians read 200 images from one centre, CHSZCAMS, in a crossover design. What can that setting tell us about how the tool would work with real patients in a clinic?
Answer in the threadIf you were one of the senior physicians whose reads agreed more with colleagues after using the AI, what would make you trust or override its Ni classification?
Answer in the thread
If you have more time
The model was trained on 5404 retrospectively collected images. What would you ask about which patients and which centres those images came from?
Answer in the threadWhat result from a study run in live clinics, with actual biopsy outcomes, would change your view of the second-reader tool?
Answer in the thread
ace
What readers say