finetune a segmentation model on a medical classification task

maybe models in this space would generalize better if we warmed em up a little, with a good image representation like DINO or BYOL or whatever. totally possible this trick has already been tried and proved/failed.