Learning to identify diseases from healthy medical images



Sue Liu


Research Data Scientist - Mudano Limited

Abstract: Using machine learning to detect diseases from medical imaging has made remarkable progress in recent years, achieving expert-level accuracy in some cases. However, nearly all the published machine learning methods are supervised, in that they need to learn from labelled data. Labelling data is an expensive operation and in the case of medical images, the expertise required for labelling may not be available. However, when the anomalous data is only a very small subset of the overall data, research has shown that it is possible to build an unsupervised machine learning model to detect anomalies. This means all the data can be used 'as is' without any labelling. This talk introduces two state-of-the-art unsupervised deep learning models, variational autoencoder (VAE) and generative adversarial network (GAN) and demonstrate that once trained on the unlabelled images, they are able to perform inference to determine whether a test image is normal or not. Experiments were carried out using the Optical Coherence Tomography dataset from Kaggle, and the results show that these models are able to detect retina diseases with comparable accuracy to supervised methods.

Bio: Dr Sue Liu is a research data scientist at Mudano, an Edinburgh based technology company specialising in machine learning and project management. She holds a PhD in mathematics from the University of Cambridge and an MSc in Medicinal Chemistry from the Open University. She works on applying machine learning to solve real-life problems in fields including finance, chemistry and computer vision. She also collaborates with the department of computer science at the University of St Andrews to carry out original research on medical imaging.

PosterSteven ScottPoster