Shuffling the VALIDATION data leads to a lower Validation Loss (U-Net like architecture)

Question

I am training a U-Net like architecture for medical image segmentation at the moment. Therefore I am using a custom loss function:

def soft_dice_coef(y_true,