Limitations of clDice for Evaluating Pancreatic Duct Segmentation in CT

Background/Objectives: Centerline Dice (clDice) emphasizes skeleton overlap, but a high score does not establish connectivity. This empirical study, which proposes no new metric, measured how often high-scoring pancreatic-duct segmentations on computed tomography (CT) differ from their references in connected-component count, and what changes that frequency. Methods: Dice, clDice and component counts were compared for 235 held-out PANORAMA CT predictions of a public model. The references were algorithm-generated, so a difference shows disagreement, not anatomical correctness. Thresholds, small-component cleanup and skeleton-based counting were varied. Further tests used a second model, known cuts, synthetic tubes and retinal labels. Synthetic-tube radii illustrate the mechanism, not clinical thresholds. Results: Among 81 predictions with both scores above 0.85, 51 (63.0%; 95% CI 52–73%) differed in component count; a second model gave 32 of 55 (58.2%). Among 86 predictions with Dice above 0.85, cleanup alone moved discordant cases between 25 and 60, and skeleton-based counting reclassified seven cases. After one known cut, 194 of 199 references kept both scores above 0.85 despite the break. Conclusions: High clDice does not ensure matching connectivity. Studies should report a direct connectivity measure alongside clDice and state their cleanup and skeletonization conventions, because these conventions alone changed the reported frequency.