In short, Dice is a very poor "metric" (it is not even a metric in mathematical sense) due to its insensitivity and shape dependency. There have been many questions about this on ResearchGate, for example have a look at here: https://www.researchgate.net/post/How_can_I_compare_a_segmented_image_to_the_ground_truth or search for Hausdorff and Dice.