Abstract
BACKGROUND: Manual cytological diagnosis for early esophageal squamous cell carcinoma (early ESCC) and high-grade intraepithelial neoplasia (HGIN) is unsatisfactory. Herein, we have introduced an artificial intelligence (AI)-assisted cytological diagnosis for such lesions. METHODS: Low-grade squamous intraepithelial lesion or worse was set as the diagnostic threshold for AI-assisted diagnosis. The performance of AI-assisted diagnosis was evaluated and compared to that of manual diagnosis. Feasibility in large-scale screening was also assessed. RESULTS: AI-assisted diagnosis for abnormal cells was superior to manual reading by presenting a higher efficiency for each slide (50.9 ± 0.8 s vs 236.8 ± 3.9 s, p = 1.52 × 10(-76) ) and a better interobserver agreement (93.27% [95% CI, 92.76%-93.74%] vs 65.29% [95% CI, 64.35%-66.22%], p = 1.03 × 10(-84) ). AI-assisted detection showed a higher diagnostic accuracy (96.89% [92.38%-98.57%] vs 72.54% [65.85%-78.35%], p = 1.42 × 10(-14) ), sensitivity (99.35% [95.92%-99.97%] vs 68.39% [60.36%-75.48%], p = 7.11 × 10(-15) ), and negative predictive value (NPV) (97.06% [82.95%-99.85%] vs 40.96% [30.46%-52.31%], p = 1.42 × 10(-14) ). Specificity and positive predictive value (PPV) were not significantly differed. AI-assisted diagnosis demonstrated a smaller proportion of participants of interest (3.73%, [79/2117] vs.12.84% [272/2117], p = 1.59 × 10(-58) ), a higher consistence between cytology and endoscopy (40.51% [32/79] vs. 12.13% [33/272], p = 1.54 × 10(-) 8), specificity (97.74% [96.98%-98.32%] vs 88.52% [87.05%-89.84%], p = 3.19 × 10(-58) ), and PPV (40.51% [29.79%-52.15%] vs 12.13% [8.61%-16.75%], p = 1.54 × 10(-8) ) in community-based screening. Sensitivity and NPV were not significantly differed. AI-assisted diagnosis as primary screening significantly reduced average cost for detecting positive cases. CONCLUSION: Our study provides a novel cytological method for detecting and screening early ESCC and HGIN.