A multicenter retrospective study from four tertiary hospitals in China found that a deep learning endoscopy model distinguished benign from malignant gastric ulcers in real time, with overall precision and recall of 0.91 and specificity of 0.95. For malignant ulcer recognition specifically, the system achieved precision 0.90, recall 0.91, specificity 0.99, and processed video at 113 frames per second with 8.84 ms latency.

Why It Matters To Your Practice

• Real-time AI support during endoscopy could help clinicians identify Malignant neoplasm within gastric ulcers more consistently.

• High specificity for malignant ulcers may help prioritize biopsies and reduce missed cancers in visually ambiguous lesions.

• Dynamic video-based assessment is more aligned with live procedural workflow than still-image tools.

Clinical Implications

• The model was built on an improved YOLOv8 architecture with an illumination attention module for instance segmentation and classification.

• Training data included 9,820 benign ulcer images, 1,727 malignant ulcer images, and 15,791 normal mucosa images.

• Performance suggests feasibility for real-time deployment as a second reader during upper endoscopy rather than a standalone diagnostic tool.

Insights

• This was a multicenter, retrospective validation study, which improves generalizability versus single-center development alone.

• The strongest performance signal was for ruling in malignant ulcers, with specificity of 0.99 in the malignant ulcer task.

• Because the study was retrospective and conducted in tertiary centers in China, prospective and external validation will be important before broad adoption.

The Bottom Line

• Deep learning can now flag malignant gastric ulcers during live endoscopy with high accuracy and minimal latency.

• For clinicians, the near-term value is likely better lesion targeting, more standardized recognition, and smarter biopsy decisions.