Limitations of clDice for Evaluating Pancreatic Duct Segmentation in CT

The centerline Dice coefficient (clDice) is used to evaluate tubular segmentations, but a high score does not directly test whether the predicted centerline is connected. We examined this limitation for structures only a few voxels wide. First, we eroded smooth and irregular synthetic tubes while varying their radius. Second, we compared Dice, clDice, ccDice, connected-component error and Betti matching on 235 held-out PANORAMA pancreatic-duct predictions. We also tested operating-point sensitivity, component cleanup, connectivity definitions, local thickness, graph measurements and two manually labeled retinal datasets. In the tested synthetic setting, mean clDice first exceeded 0.90 at a radius of 2.50 voxels, whereas every tested skeleton remained connected only from 2.90 voxels. At radius 2.80, mean clDice was 0.990 although none of the five skeletons preserved the reference component count. In the CT cohort, Dice and clDice were strongly associated across cases (Spearman ρ=0.857), yet 54 of 86 predictions with Dice >0.85 differed from the reference in component count; 51 of these also had clDice >0.85. The 2.90-voxel value is specific to the synthetic setup and is not transferred to clinical CT. For structures only a few voxels wide, clDice should be reported together with a direct connectivity measure.