Machine learning’s success may depend on addressing 'gray areas' of cancer diagnosis