Improving the reliability of neural networks for medical imaging

点击翻译按钮获取中文摘要

Over the past decade, large strides have been made in machine learning and artificial intelligence for analysing medical imaging data. These advances have been primarily fuelled by development of advanced deep neural networks. However, a major problem with this technology is that these models are not robust, and their performance can be influenced by variations of the data characteristics. Specifically, it has been commonly observed that performance of neural networks degrades significantly when they are applied on data that present differences from the data that were used for training the models. For example, deep neural networks have been shown to perform poorly on data from people from ethnic minorities that were under-represented in the training data. This constitutes a major barrier to implementing such models in high-risk applications such as processing medical images - where an incorrect diagnosis could have major consequences. Therefore, creating methods to improve the robustness and reliability of neural networks will allow safe and effective integration of this technology in medical applications, which promises to deliver significant benefits to healthcare such as reducing the staff strain, accelerating and improving diagnosis of disease, and ultimately improving patient outcome.The aim of this project is to work on improving the reliability of neural networks for the study of medical images - such as X-ray, CT and MRI images. One aim of the project is to create methods to enable neural network models to quantify their own uncertainty in their predictions. This will provide them the ability to inform the user when it is likely that a prediction may be wrong (known as 'failing gracefully'). This would allow the user, such as clinicians, to judge whether to include the model's prediction in follow-up decision making or discard the model's output. Moreover, another aim is to create a tool for the neural network to clearly explain its reasoning for why it made a prediction based upon the data it was given. Using techniques such as causal machine learning, models will be able to explain what features led to a prediction, bolstering their interpretability, as well as explaining what input data will cause failures. For example, many baseline models currently misidentify motion artefacts in images as pathology, which could be studied using causal machine learning.This research aims to enable safer and more reliable use of deep neural networks for healthcare. This can have major impact as the technology has the potential to improve diagnosis and treatment, but its integration in clinical workflows is currently hindered due to the discussed technology limitations. Furthermore, advancements from this research will also facilitate the implementation of deep neural networks in other high-risk applications beyond healthcare - such as self-driving vehicles.This project is multi-disciplinary, combining advances in computer science with medical data. Therefore, this project falls within the EPSRC Healthcare Technologies theme.