Producing radiologist-quality reports for interpretable artificial intelligence

Current approaches to explaining the decisions of deep learning systems for\nmedical tasks have focused on visualising the elements that have contributed to\neach decision. We argue that such approaches are not enough to "open the black\nbox" of medical decision making systems because they are missing a key\ncomponent that has been used as a standard communication tool between doctors\nfor centuries: language. We propose a model-agnostic interpretability method\nthat involves training a simple recurrent neural network model to produce\ndescriptive sentences to clarify the decision of deep learning classifiers.\n We test our method on the task of detecting hip fractures from frontal pelvic\nx-rays. This process requires minimal additional labelling despite producing\ntext containing elements that the original deep learning classification model\nwas not specifically trained to detect.\n The experimental results show that: 1) the sentences produced by our method\nconsistently contain the desired information, 2) the generated sentences are\npreferred by doctors compared to current tools that create saliency maps, and\n3) the combination of visualisations and generated text is better than either\nalone.\n