When considering diagnoses from imaging exams, radiologists rely on prior knowledge accumulated through experience in their practice but also patient specific background information. However, the amount of information potentially available – e.g. from clinical history, prior imaging– is ever increasing, resulting in a risk of information overload.
A well-integrated decision support system can aid the process upon which a diagnostic decision is rendered. Such systems can assist with specific tasks such as delineating pathology, and help with the consideration of differential diagnoses by retrieving similar cases from an institution’s database.
Central to this goal is the definition of appropriate data representation spaces that relate to clinical similarity that will empower the underlying decision support mechanisms.
In collaboration with Toshiba Medical Visualization Systems, this project will develop methods, algorithms and theory that realize decision support systems via deep representation learning mechanisms based on both imaging and non-imaging clinical data. This requires clinically meaningful statistical distances to be learned, taking advantage of both annotated and weakly (or non) annotated data – the latter being plentiful and easier to source. Such semi-supervised approaches to learning are increasingly sought-after in many other fields (e.g. in multimedia where both video, audio and text are available). Of particular interest is how one can balance the contributions of different information sources (imaging vs. non-imaging sources) and how to deal with missing or incomplete observations. The project will draw inspiration from generative methods to build robustness and unearth the importance of each source. For the system to be credible it should be able to present its reasoning by identifying the most significant information in each case, and exposing its degree of confidence. Consequently, this project will also consider how generative methods can provide confidence estimates.
To achieve this, we are looking for an enthusiastic and strongly motivated student to join our team. He/she will have the opportunity to collaborate with Toshiba researchers in Edinburgh and other collaborators throughout the world.
Dr Sotirios Tsaftaris
Candidates should have a Master’s level education (or in exceptional cases an excellent (upper second class and higher) Bachelor’s degree) in electronic/electrical engineering, computer science (informatics), physics or closely related subjects. Any prior experience in medical image analysis is desirable but not necessary. Prior exposure to machine learning is desirable but not essential. The candidate will be expected to have a high level of analytical and investigative skills. The candidate should have good programming skills (e.g., Python, Matlab) and a solid mathematical background.
Further information on English language requirements for EU/Overseas applicants.
This position can be fully funded for 42 months (3.5 years) and is open to UK nationals and EU nationals (if they have resided for more than 3 years in the UK).