AN IMAGE-BASED MODEL FOR ASSISTING IN DIAGNOSING MALIGNANT ESOPHAGEAL LESIONS DURING LUGOL'S CHROMOENDOSCOPIC EXAMINATION.

Abstract

Introduction: Image-based diagnostic tools that aid endoscopists to biopsy putative esophageal malignant lesions are essential for ensuring the standardization and quality of Lugol's chromoendoscopy. But there's no such model available yet.

Methods: We developed a diagnostic model using endoscopic Lugol-unstained lesions (LULs) features and baseline data from 1099 individuals enrolled from a large-scale population-based ESCC screening cohort. 603 participants from a clinical outpatient cohort were included as the external validation set. High-grade intraepithelial neoplasia and above (HGINA) lesions identified at baseline or within 1 year after screening were defined as outcome. The final model was determined using logistic regression analysis by the Akaike Information Criterion.

Results: The optimal diagnostic model contained the size, irregularity, sharp border of LUL, age and body mass index of the participant, with the area under the curve of 0.83 (95% CI: 0.78-0.87) in the development set, 0.81 (95% CI: 0.77-0.86) in the internal validation set, and 0.87 (95% CI: 0.84-0.90) in the external set. This model stratified individuals with LULs into low-, moderate-, and high-risk groups based on tertiles of predicted probabilities. The high-risk group accounted for < 40% participants but enriched 80.8% and 82.7% of HGINA cases in the development and external validation sets, respectively, achieving detection ratios 16.2 and 11.0 times higher than the low-risk group.

Discussion: Our model can help maintain consistency and accuracy in detecting esophageal malignancy through Lugol's chromoendoscopy, particularly in primary healthcare units in high-risk rural areas.