Learning from limited data : three case studies in healthcare

Abstract

Medical data are generally limited in supply due to reasons of confidentiality, ethics and expense. This poses a challenge in applying machine learning techniques to medical problems. Depending on what data are available, different techniques might be considered. In this thesis we examine different approaches for application to fall detection, stroke signs detection and medical image retrieval. For the problem of detecting falls from accelerometer data we consider novelty-based detection approaches. It is difficult to acquire examples of falls from the target population of elderly people. However, it is possible to gather sufficient numbers of examples from their normal daily activities for training of a normality model. We compare that approach with supervised methods trained on simulated falls performed by young volunteers and we find that novelty detection approaches are a promising method when applied to this low dimensional problem. The problem of detecting stroke signs in Non-Contrast Computer Tomography (NCCT) is very challenging due to the subtlety of the intensity and texture changes in the case of ischaemic regions and the proximity of bone in the case of dense vessel signs. Furthermore, obtaining manual segmentation of stroke signs is a very laborious and expensive process. Inspired by the workflow of neuroradiologists during their reading of NCCT we incorporate the appearance of the contralateral anatomy and the atlas-encoded spatial location into a Convolutional Neural Network (CNN) architecture. We find that incorporation of domain knowledge helps to achieve better stroke sign detection performance than a standard CNN when trained on a limited number of annotated datasets. Finally, we conduct a feasibility study in which we explore the types of labels required fort raining of an embedding model for matching similar images. It is hard to foresee all possible variations of pathologies that a patient may present with in the hospital, therefore a similar image retrieval system which can cope with pathologies unseen during training may offer a more generalisable solution than a dedicated pathology classifier. We find that training the model on multiple tasks helps to achieve a more useful compressed representation.