In the realm of knee pathology, magnetic resonance imaging (MRI) has the advantage of visualising all structures within the knee joint, which makes it a valuable tool for increasing diagnostic accuracy. We previously developed KneeTex (Spasić et al., 2015), a system that extracts clinical observations from MRI reports. As a result, MRI reports are formally structured and coded, which improves accessibility of information contained in these documents.
However, the system stands to be improved in two ways. First, it does not differentiate between spatial relations that may exist between clinical findings and anatomical sites. It simply detects that two entities are related, but does not assert in which way (e.g. in, on, near), thus important clinical information gets lost. The first goal of this project would be to extend KneeTex functionality by implementing a method that automatically extracts and classifies different types of spatial relations.
Originally, KneeTex was implemented as rule-based system, which relies on a set of highly sophisticated hand-crafted rules to extract free-text information and convert it into a pre-defined structure. Unfortunately, the rule-based approach does not generalise well to other domains (e.g. cancer) or document types (e.g. X-ray reports), which highlights the second aspect of the system that needs to be improved. Traditional machine learning approaches also require a lot of manual feature engineering to perform well unlike deep learning, which discovers the most useful features from the training data automatically. The second goal of the project would be to re-implement KneeTex functionality using a deep learning approach, which has proven potential to make sense of large amounts of text data. The original system, which performs at human-like level, will serve as the baseline to test a widespread assumption that deep learning performs as well as humans.
The third goal of the project would be to investigate how portable the deep learning approach is by applying it to other domains, e.g. extracting cancer-related information from X-ray reports (Spasić et al., 2014).
Irena Spasić, Bo Zhao, Christopher Jones, Kate Button (2015) KneeTex: An ontology-driven system for information extraction from MRI reports. Journal of Biomedical Semantics, Vol. 6, 34 [DOI: 10.1186/s13326-015-0033-1]
EPSRC Doctoral Training Partnership (DTP) studentship, providing full UK/EU tuition fees (£4,121 per annum in 2016/17) and doctoral stipend matching UK Research Council National Minimum (£14,296 per annum in 2016/17). Funding is 50% EPSRC DTP and 50% Cardiff University. UK Research Council eligibility conditions apply. Applicants for a studentship must have obtained, or be about to obtain, a 2.1 degree or higher in a relevant subject or a masters degree with distinction in the research dissertation in a relevant discipline. Full details and How to Apply: http://www.cardiff.ac.uk/study/postgraduate/funding-and-fees/funding-options/research-councils/epsrc-studentships
How good is research at Cardiff University in Computer Science and Informatics?
Research output data provided by the Research Excellence Framework (REF)