AIM and Explainable Methods in Medical Imaging and Diagnostics

Abstract

Medical imaging and diagnostics have benefited from recent advances in machine learning in general and deep learning in particular. There are a large number of studies that report significant performance in a variety of medical image analysis tasks. At the same time, advances in time series analysis of electronic health records and physiological signals from a variety of sources (including clinical and nonclinical wearable sensors) have enabled efficient clinical pipelines and diagnostic performance. While there are a large number of studies that report such advances and use cases, there is a certain level of resistance to clinical adaptability of such methods. We attribute one of the major reasons for this to be the black-box nature of such machine learning and data-driven models. While clinical decision-making, for example, using radiology, relies not only on radiographic information but also on clinical expertise (that could rely on clinical history of a patient as well as demographic information), such informed decision makes the backbone of clinical practice. Hence, machine learning should adapt to making such informed decisions for wider adaptability and finding more practical use cases. We argue that explainable methods, where a machine learning-based model provides supporting information for their decisions making them understandable, will be more representative of what is required in clinical practice. In turn, such methods would augment the current medical imaging and diagnosis pipeline and hence allow adaptation of artificial intelligence in clinical workflows. This would transform the healthcare industry with far-reaching benefits for both clinical practitioners and patients enabling precise and personalized medicine.