AI Helps in Classification of the Z-Line During Endoscopy

The squamocolumnar junction (SCJ), or the Z-line, is a morphological marker in the esophagus that assists in diagnosing various esophageal conditions, including Barrett’s esophagus. A regular Z-line has a smooth, concentric appearance, whereas an irregular Z-line is jagged and may extend less than 1 cm above the gastroesophageal junction. Irregular Z-lines are common, yet there has been controversy regarding whether to obtain biopsy specimens from them.

This study utilized 849 high-quality Z-line images to train a deep-learning model for automated segmentation of the SCJ. A second dataset of 58 videos was then used to test the model’s performance compared with 10 gastroenterologists (5 esophageal experts) who rated each video as regular or irregular. There was fair agreement (Fleiss’ kappa [FK] =0.39) among all gastroenterologists when rating the Z-line, with moderate agreement among esophageal experts (FK=0.42). The computer-assisted model showed an accuracy of 78% in distinguishing regular from irregular Z-lines (wavelet energy coefficient cutoff of 1.53 x10⁷).