TOPLINE

A meta-analysis of 14 studies showed that endoscopy-based deep learning algorithms outperformed junior endoscopists for detecting early esophageal squamous cell carcinoma, performed comparably to senior endoscopists, and improved endoscopists’ performance when used as assistance, especially among less experienced specialists.

METHODOLOGY

• Traditional endoscopic methods such as white light imaging and image-enhanced endoscopy have important limitations in detecting early esophageal squamous cell carcinoma, whereas deep learning offers more objective and consistent recognition of lesions that can better identify subtle abnormalities and support endoscopists in clinical practice.
• Researchers conducted a systematic review and meta-analysis searching popular databases through March 20, 2025, with an update search through March 1, 2026, to compare the diagnostic performance of endoscopy-based deep learning algorithms with that of endoscopists with different experience levels for detecting early esophageal squamous cell carcinoma.
• A total of 14 studies that evaluated endoscopy-based deep learning models for the detection of early esophageal squamous cell carcinoma against pathologic biopsy as the reference standard met the inclusion criteria.
• The analysis included 18 deep learning study comparisons, nine comparisons of general endoscopists, 11 comparisons of junior endoscopists (physicians with < 5 years of experience who completed 1000-2000 upper gastrointestinal endoscopies), and 11 comparisons of senior endoscopists (physicians with > 10 years of experience or those who completed more than 10,000 procedures).
• The primary outcomes were sensitivity, specificity, diagnostic odds ratio, and area under the receiver operating characteristic curve for deep learning algorithms and endoscopists with varying experience levels; these parameters measured how well they detected early esophageal squamous cell carcinoma and distinguished it from nondiseased tissue.

TAKEAWAY

• Deep learning algorithms demonstrated a sensitivity of 0.94, specificity of 0.88, area under the receiver operating characteristic curve of 0.96, and diagnostic odds ratio of 106.76 outperforming all endoscopists, who showed a sensitivity of 0.82, specificity of 0.78, area under the receiver operating characteristic curve of 0.87, and diagnostic odds ratio of 16.71 (P < .050 for all).
• Deep learning algorithms achieved higher sensitivity (0.94 vs 0.76; P < .001) and area under the receiver operating characteristic curve (0.96 vs 0.84; P < .001) than junior endoscopists, but no significant differences were found between deep learning algorithms and senior endoscopists across all metrics (P > .050).
• With the assistance of deep learning, junior endoscopists showed increased area under the receiver operating characteristic curve from 0.85 to 0.94 (P < .001), whereas senior endoscopists demonstrated improvements in sensitivity, specificity, and area under the receiver operating characteristic curve, although the differences did not reach statistical significance.
• Subgroup analyses showed no significant performance difference by lesion size; sensitivity was higher for lesions in the cervical-upper thoracic esophagus than in the middle-lower thoracic or abdominal esophagus, area under the receiver operating characteristic curve was higher with white light imaging than with image-enhanced endoscopy, and performance was lower in lesion-based analyses than in image-based or patient-based analyses.

IN PRACTICE

“In conclusion, this meta-analysis reveals endoscopy-based DL [deep learning] demonstrates consistently high diagnostic accuracy for early ESCC [esophageal squamous cell carcinoma] detection, outperforming junior endoscopists across all metrics. AI assistance benefits both junior and senior endoscopists. However, current evidence does not suggest DL superiority over experienced senior endoscopists,” the authors wrote.

SOURCE

The study was led by Xiaoyan Men, Department of Endoscopy, The NO.4 People’s Hospital of Hengshui, Hebei, China. It was published online in the American Journal of Gastroenterology.

LIMITATIONS

The study is limited by the fact that most included studies were single-center and retrospective. In addition, the patient populations were largely homogeneous and drawn from Asian cohorts, which limits generalizability. Differences in deep learning architectures and validation methods also make comparisons across studies more difficult.

DISCLOSURES

The authors reported that no funding was received for this study. No relevant conflicts of interest or commercial interests were disclosed by the authors.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.