Hi @rbell and welcome to NeuroStars!
What kind of T1w image are you using? 7T data? You may want to provide your own brain mask as it looks like the brain mask calculated by fmriprep on these datasets is cutting out the cerebellum. You could try to give your own skull stripped image and use --skull-strip-t1w skip in your fmriprep call. I am not entirely sure that this will improve the anatomical segmentation.
Otherwise, you could follow what is said in this recent thread: fMRIPrep Normalization failed in a problematic subject