Keynote: Beyond Supervised Learning

Abstract

Machine learning has been widely regarded as a solution for diagnostic automation in medical image analysis, but there are still unsolved problems in robust modelling of normal appearance and identification of features pointing into the long tail of population data. In this talk, I will explore the fitness of machine learning for applications at the front line of care and high throughput population health screening, specifically in prenatal health screening with ultrasound and MRI, cardiac imaging, and bedside diagnosis of deep vein thrombosis. I will discuss the requirements for such applications and how quality control can be achieved through robust estimation of algorithmic uncertainties and automatic robust modelling of expected anatomical structures. I will also explore the potential for improving models through active learning and the accuracy of nonexpert labelling workforces. However, I will argue that supervised machine learning might not be fit for purpose, as it cannot handle the unknown and requires a lot of annotated examples from well-defined pathological appearance. This categorization paradigm cannot be deployed earlier in the diagnostic pathway or for health screening, where a growing number of potentially hundred-thousands of medically catalogued illnesses may be relevant for diagnosis. Therefore, I introduce the idea of normative representation learning as a new machine learning paradigm for medical imaging. This paradigm can provide patient-specific computational tools for robust confirmation of normality, image quality control, health screening, and prevention of disease before onset. I will present novel deep learning approaches that can learn without manual labels from healthy patient data only. Our initial success with single class learning and self-supervised learning will be discussed, along with an outlook into the future with causal machine learning methods and the potential of advanced generative models [1].