Would you prefer high dice score or high accuracy?

Hi all,
This is more for my curiosity.
I’m training a UNET model from scratch. I see it can reach 92% pixelwise accuracy while reaching 0.72 dice score.
I noticed that sometime accuracy can be high but dice score is lower than the best dice score.
Would you prefer to save the model when accuracy get higher or when it’s the dice score that rises?