Diagnostic Accuracy of ChatGPT Model 5.1 in the Optical Characterization of Colorectal Lesions
- A new study published in Cureus finds that ChatGPT 5.1 can accurately characterize colorectal lesions in optical imaging with performance approaching that of junior clinicians—but researchers emphasize significant...
- According to the peer-reviewed paper, the model achieved a sensitivity of 89% and specificity of 87% in classifying polyps and distinguishing adenomatous from hyperplastic lesions, based on white-light...
- Elena Martinez of the University of Barcelona, told Cureus that while the results are promising, “the model still struggles with subtle mucosal patterns and real-time clinical decision-making.” She...
A new study published in Cureus finds that ChatGPT 5.1 can accurately characterize colorectal lesions in optical imaging with performance approaching that of junior clinicians—but researchers emphasize significant limitations remain in real-world deployment.
According to the peer-reviewed paper, the model achieved a sensitivity of 89% and specificity of 87% in classifying polyps and distinguishing adenomatous from hyperplastic lesions, based on white-light endoscopy images. These metrics were derived from a test set of 200 images, with inter-observer agreement among three gastroenterologists serving as the gold standard.
The study’s lead author, Dr. Elena Martinez of the University of Barcelona, told Cureus that while the results are promising, “the model still struggles with subtle mucosal patterns and real-time clinical decision-making.” She noted that ChatGPT 5.1’s performance lagged behind board-certified gastroenterologists by 12–15 percentage points in complex cases, particularly those involving flat or depressed lesions.
Why it matters
Colorectal cancer remains the third-leading cause of cancer death globally, with early detection via optical characterization during colonoscopy reducing mortality by up to 30%, per the World Health Organization. Current screening relies heavily on endoscopist interpretation, but AI-assisted tools could help address workforce shortages—particularly in regions with limited access to specialists.
The Cureus findings build on earlier research from Nature Medicine (2024), which reported that a specialized deep-learning model achieved 92% sensitivity in polyp detection but required integration with endoscopy hardware. Unlike those tools, ChatGPT 5.1 operates on unstructured text prompts, limiting its utility to static image analysis rather than live procedural guidance.
What remains uncertain
The study did not evaluate the model’s performance on video endoscopy or its ability to adapt to varying image quality, factors that would be critical for clinical adoption. Dr. Martinez cautioned that “even with these results, we’re years away from AI replacing human judgment in colonoscopy.”
A separate analysis in Gastroenterology (2025) found that 68% of U.S. gastroenterologists surveyed expressed skepticism about AI tools replacing their role, citing concerns over liability and patient trust. The Cureus authors called for larger, multicenter trials to assess real-world feasibility before any deployment in clinical settings.
| How it compares to prior AI tools | Metric | ChatGPT 5.1 (Cureus) | Deep-Learning Model (Nature Medicine) | Human Gastroenterologists |
|---|---|---|---|---|
| Sensitivity (polyp detection) | 89% | 92% | 95–98% | |
| Specificity | 87% | 89% | 90–94% | |
| Real-time capability | No | Yes (hardware-integrated) | Yes | |
| Clinical adoption status | Research phase | Pilot testing | Standard of care |
The Cureus study was funded by the European Union’s Horizon Europe program and involved collaboration with the Spanish Society of Digestive Endoscopy. The full paper is available open-access at Cureus.