70/100

Moderate evidence

Cite
More
EmailLinkedInX

Colon polyp AI matched senior endoscopists in a 149-image test, retrospective study finds

Study title: Deep learning for predicting high-grade dysplasia in colorectal polyps using white-light endoscopy

Evidence type
Controlled or validated comparison
Sample
n = 1,071
Sources
1 sourcePeer-reviewed
Published
3 min read

Study title: Deep learning for predicting high-grade dysplasia in colorectal polyps using white-light endoscopy

Researchers at three tertiary hospitals in China built a MobileNetV2 deep learning model that predicts high-grade dysplasia in colorectal polyps from one routine white-light colonoscopy image per lesion. It scored an AUC of 0.879 on an internal test set of 146 polyps and 0.814 and 0.803 in two outside hospitals' cohorts of 159 and 182 polyps.

Read the study abstract

BACKGROUND Accurate identification of high-grade dysplasia (HGD) in colorectal polyps is important for optimizing endoscopic management and surveillance strategies. Although deep learning has shown potential for colorectal polyp characterization, the generalizability of white-light imaging (WLI)-based models across independent clinical settings remains insufficiently established. AIM To develop and externally validate a deep learning model based on routine WLI images for predicting HGD in colorectal polyps, to e…

The gain

In a retrospective study at three Chinese hospitals, an AI model flagged high-grade dysplasia in colorectal polyp photos with AUCs of 0.814 and 0.803 in two outside cohorts, matching senior endoscopists and beating junior ones on 149 images.

In brief

In a blinded reader study of 149 images, two senior and two junior endoscopists made binary calls on the same pictures. The model performed within the range of the seniors and was more accurate overall than the juniors, which suggests it could support less experienced clinicians. A logistic regression model built from age, sex, lesion size, location and shape added no clear predictive value beyond the image-based prediction.

The results are early. The outside cohorts held only 20 and 26 high-grade cases, so the estimates are imprecise, and positive predictive values were 34.1% and 39.6%. The model was tested on static, hand-picked images, not live colonoscopy, and the authors say prospective real-time validation is needed before any clinical use.

Main points

  1. The internal test set of 146 polyps gave an AUC of 0.879 (95% CI 0.795-0.952), with 82.35% sensitivity and 82.17% specificity.
  2. Outside cohort A (159 polyps, 20 high-grade) gave an AUC of 0.814 with 75.0% sensitivity and 79.1% specificity; cohort B (182 polyps, 26 high-grade) gave 0.803 with 80.8% and 79.5%.
  3. Negative predictive value was 95.7% and 96.1% in the two outside cohorts, but positive predictive value was only 34.1% and 39.6%, meaning many false alarms.
  4. In the reader study of 149 images (32 high-grade, 21.5%), the model performed within the range of two senior endoscopists and had higher overall accuracy than two junior endoscopists.
  5. The model's five-class accuracy for polyp subtype was 52.1%, much weaker than its binary high-grade call.
  6. A clinical model using age, sex, lesion size, location and shape added no clear value beyond the image model; age, size and distal location were still independently linked to high-grade dysplasia.
  7. Calibration Brier scores were 0.092, 0.081 and 0.082 across the three cohorts, and subgroup AUCs by lesion size and location ranged from 0.74 to 0.88.

The full story

The model was trained at the Second Hospital of Jilin University on 730 patients with confirmed polyps (79 with high-grade dysplasia, 10.8%) collected from January 2024 to January 2026, split 70/10/20 at the patient level. Two outside cohorts from Jilin Provincial People's Hospital and the First Affiliated Hospital of Jinzhou Medical University, collected June to December 2025, were not used for training, threshold setting or model selection.

Each polyp contributed a single high-quality white-light image chosen by an experienced endoscopist, resized to 224 x 224 pixels. Pathology reports from each centre were the reference standard, and a central review of 149 cases agreed with the original diagnosis in 95.97% (143 of 149), with a kappa of 0.881.

In the blinded reader study, four board-certified endoscopists (two with at least 10 years' experience and over 5,000 colonoscopies, two with under 3 years) judged the same 149 images as high-grade or not. Because they gave yes/no answers, the authors did not compare AUCs directly with the model, and they call the reader study exploratory.

The authors list limitations: retrospective design with possible selection bias, few high-grade cases in the outside cohorts, validation only on white-light images, static images instead of real-time video, three tertiary Chinese hospitals only, and one image per lesion. They say the 0.2381 threshold may need local recalibration and that the findings do not establish readiness for clinical deployment.

Sources

1 on file
  1. Peer-reviewed2027-01-21

Talk it over

Three questions to start on, two more if you have time
The model's positive predictive values were 34.1% and 39.6% in the two outside cohorts, while negative predictive values were 95.7% and 96.1%. How would you weigh the missed cases it avoids against the false alarms it creates?

Start with a question

Colon polyp AI matched senior endoscopists in a 149-image test, retrospective study findshttps://truvace.com/item/deep-learning-for-predicting-high-grade-dysplasia-in-colorectal-polyps-using-whi

  1. The model's positive predictive values were 34.1% and 39.6% in the two outside cohorts, while negative predictive values were 95.7% and 96.1%. How would you weigh the missed cases it avoids against the false alarms it creates?

    Answer in the thread
  2. The outside cohorts held only 20 and 26 high-grade cases, and each polyp contributed one hand-picked static image. How much should those choices limit what the AUCs of 0.814 and 0.803 tell us?

    Answer in the thread
  3. The reader study had two senior and two junior endoscopists judging 149 images. If you were a junior endoscopist in it, how would you want to use the model's call alongside your own?

    Answer in the thread
If you have more time
  • The authors say prospective real-time validation is needed. What would a live colonoscopy test need to measure before you would trust the model's high-grade calls?

    Answer in the thread
  • All three hospitals are tertiary hospitals in China, and the authors say the 0.2381 threshold may need local recalibration. Who would you ask at a hospital you know to judge whether it fits there?

    Answer in the thread

What readers say