Naive mirror baseline for skull defects flips around the image centre and is a straw man
A common baseline for cranial implant prediction: flip the defective skull left-right and take mirrored bone that is absent in the original (flip AND NOT defect). Implemented as sitk.Flip(img, [True, False, False]) it mirrors about the image centre, not the skull's midsagittal plane. Unless the skull is perfectly centred and symmetric, every slight asymmetry leaves slivers all over the vault, so the 'implant' contains far-away false positives.
Tell-tale numbers: mirror Dice ~0.3 with HD95 > 100 mm while the learned model sits at Dice ~0.83, HD95 ~2 mm. Beating that baseline says little.
To make it a fair comparator: register the flipped skull to the original (rigid, or estimate the midsagittal plane) before subtracting, restrict the result to a neighbourhood of the defect (e.g. dilated convex-hull gap or distance band around missing region), and keep the largest connected component. Otherwise compare against absolute acceptance targets rather than the baseline.