General Profile:

Before my studies at the University of Bristol I worked in the NHS for 10 years as a diagnostic radiographer where I specialised in magnetic resonance imaging (MRI). During my time working in hospitals, I observed the implementation of new technologies within the field; the advancements most interesting to me were artificial intelligence related such as medical image interpretation through AI techniques.
My desire to understand more about AI led me to an MSc in Computer Science at the University of Bristol which has provided a solid foundation in Software Engineering. My dissertation focused on AI, researching the field of multi-agent systems by formalising the semantics of a negotiation protocol to help capture accountability between agent interactions.
From my career in the NHS I already have an interest in the applications of human centred AI in the healthcare field, however my dissertation has broadened my horizons to look outside of this scope. I joined the Interactive AI CDT because I want to explore more AI-related research fields before commencing my PhD thesis.

PhD Project Summary:

Due to the growing and aging population, there is an increased demand on imaging services within the NHS. It is important that patients are seen in a timely manner, but current NHS staff shortages are causing increased waiting lists, with the Royal College of Radiologists reporting that this is affecting patient safety. As a result, the need for automated methods for medical image analysis is growing increasingly important. Recent developments in Deep Learning have achieved state-of-the-art results for most automated medical image analysis tasks. Despite these improvements, the gold standard for most tasks is still achieved by medical professionals, and as such research into the area is still ongoing.

The automatic generation of radiology reports seeks to create a free text description for a given clinical examination. This PhD project will be focused around researching and creating models for this task, which will be achieved through combining computer vision (CV) and natural language processing (NLP) techniques. At its simplest level, CV models will be utilised to extract the relevant attributes from an image or volume while NLP techniques will be utilised to generate sentences from the extracted attributes.

Most methods within automated medical image analysis make their predictions based on an individual radiology examination. A multimodal approach, considering text based clinical indications and other associated clinical data is thought to create a more accurate prediction and better reflects how medical professionals naturally operate.

The main aim of this research is to achieve automated radiology report generation utilising multimodal data. This work will be completed through building upon previous work within the domain, where reports have been generated from a 2D clinical image. To our knowledge, there has not been much work on multi-modal report generation, nor has there been any research utilising 3D volumes of imaging data such as CT or MRI scans. The project will initially focus on understanding the feasibility of multi-modal report generation through researching techniques that could make this task achievable.

Other than the multimodal problems complexity, there are other challenges to overcome with this work. These include class imbalance within medical images, where the pathologies to be found are small in comparison to the overall image. The relative size of medical datasets is also an issue, as their size is often lower than in other domains due to privacy concerns. Solutions to these technical hurdles will be investigated during the project.