Towards an AI diagnosis like the doctor's: How can we make 'lazy' artificial intelligence more transparent and relevant to the clinic?