How about I create the positive/negative diagnosis images with a human+stable diffusion, and use that for training my classifier?