Endoscopy-based Deep Learning Algorithms versus Endoscopists in Early Esophageal Squamous Cell Carcinoma Detection: A Systematic Review and Meta-Analysis.

Abstract

To compare the diagnostic performance of endoscopy-based deep learning (DL) algorithms with endoscopists of different experience levels for detecting early esophageal squamous cell carcinoma (ESCC). PubMed, Embase, and Web of Science were searched to March 20, 2025, with an update to March 1, 2026. Eligible studies evaluated endoscopy-based DL models for early ESCC against pathological biopsy. Pooled sensitivity, specificity, and area under the curve (AUC) were estimated with a bivariate random-effects model. From 1,604 records initially identified, 14 studies met the inclusion criteria. Overall, DL outperformed all endoscopists in sensitivity (0.94 vs. 0.82; P = 0.01), specificity (0.88 vs. 0.78; P = 0.02), AUC (0.96 vs. 0.87; P < 0.001), and DOR (106.76 vs. 16.71; P = 0.03), particularly against juniors. With DL assistance, junior endoscopists achieved a higher AUC (0.94 vs. 0.85; P < 0.001); seniors also improved in sensitivity (0.97 vs. 0.91), specificity (0.94 vs. 0.90), and AUC (0.97 vs. 0.95), though not significantly (P = 0.10, 0.06, 0.11). Performance did not differ by lesion size (<2 cm vs. ≥2 cm), but sensitivity varied by location (Ce-Ut vs. Mt/Lt-Ae: 1.00 vs. 0.78; P = 0.01). White-light imaging-based DL yielded a higher AUC than image-enhanced endoscopy-based DL (0.98 vs. 0.96; P = 0.02). Endoscopy-based DL shows high accuracy for early ESCC detection and appears to outperform endoscopists overall, especially juniors, while also enhancing both groups' performance. Current evidence does not demonstrate DL surpassing senior endoscopists. Prospective, multicenter studies across diverse populations are warranted.