Radiology: Artificial Intelligence will host its second tweet chat on July 1, 2020, from 8:00 to 9:00 pm Eastern Daylight Time (U.S.), on the topic of interpretability of AI algorithms in radiology. The tweet chat will be moderated by Dr. Despina Kontos, deputy editor of this journal and associate professor of radiology at the University of Pennsylvania, and Dr. Aimilia Gastounioti, a research associate in the department of radiology at the University of Pennsylvania. The article discusses the methods that allow AI systems to explain their decisions through visualization, counterexamples, and semantics--and the many challenges to bring interpretability methods into clinical practice. Enhancing interpretability is essential to allow for AI systems to be trusted and verified for faster and more reliable adoption into clinical workflows. They explore whether patients and radiologists can better trust a model that explains its decisions, and how interpretability may accelerate the translation of deep learning tools into clinical practice.
This is a condensed summary of an international multisociety statement on ethics of artificial intelligence (AI) in radiology produced by the ACR, European Society of Radiology, RSNA, Society for Imaging Informatics in Medicine, European Society of Medical Imaging Informatics, Canadian Association of Radiologists, and American Association of Physicists in Medicine. AI has great potential to increase efficiency and accuracy throughout radiology, but it also carries inherent pitfalls and biases. Widespread use of AI-based intelligent and autonomous systems in radiology can increase the risk of systemic errors with high consequence and highlights complex ethical and societal issues. Currently, there is little experience using AI for patient care in diverse clinical settings. Extensive research is needed to understand how to best deploy AI in clinical practice.
Radiology is one of the most essential fields in clinical medicine. Experts in this field are specialists in deciphering and diagnosing disease based on various imaging modalities, ranging from ultrasound, magnetic resonance imaging (MRI), computerized tomography (CT), and x-rays. Studies have shown that the use of radiology in clinical practice has exponentially grown over the years: at the Mayo Clinic, between the years 1999 to 2010, use of CT scans increased by 68%, MRI use increased by 85%, and overall use of imaging modalities for diagnostic purposes increased by 75%, all numbers that have likely continued to rise, and indicate the sheer demand and growth of this robust field. A unique proposal that has become prominent over the last few years to help alleviate this increased demand is the introduction of artificial intelligence (AI) technology into this field. Simply put, the premise of AI as an addition to the practice of radiology is straightforward, and has been envisioned in two main ways: 1) a system that can be programmed with pre-defined criteria and algorithms by expert radiologists, which can then be applied to new, straightforward clinical situations, or 2) deep learning methods, where the AI system relies on complex machine learning and uses neural-type networks to learn patterns via large volumes of data and previous encounters; this can then be used to interpret even the most complicated and abstract images.
I first learned about artificial intelligence (AI) in healthcare during my undergraduate studies. As a biomedical engineering major planning a career in medicine, the concepts of AI in healthcare intrigued me. There was talk that these innovations could revolutionize medical practice in a way that would disrupt traditional physician roles. When I entered medical school, I expected to see these innovations translated into medical practice. The launch of Watson Health shortly after my medical school graduation instilled hope that these innovations would finally make it into my medical practice during residency training.
Current approaches to explaining the decisions of deep learning systems for medical tasks have focused on visualising the elements that have contributed to each decision. We argue that such approaches are not enough to "open the black box" of medical decision making systems because they are missing a key component that has been used as a standard communication tool between doctors for centuries: language. We propose a model-agnostic interpretability method that involves training a simple recurrent neural network model to produce descriptive sentences to clarify the decision of deep learning classifiers. We test our method on the task of detecting hip fractures from frontal pelvic x-rays. This process requires minimal additional labelling despite producing text containing elements that the original deep learning classification model was not specifically trained to detect. The experimental results show that: 1) the sentences produced by our method consistently contain the desired information, 2) the generated sentences are preferred by doctors compared to current tools that create saliency maps, and 3) the combination of visualisations and generated text is better than either alone.