Rethinking Transfer Learning for Medical Image Classification