Project Description

Modern machine learning approaches for image analysis

Our project aims to develop a deep learning (DL) artificial intelligence method for assessing CT brain imaging in patients with acute ischaemic stroke. We have used over 5000 CT brain image sets and associated highly characterised clinical metadata from a large completed randomised controlled trial in stroke (the Third International Stroke Trial, IST-3) to develop, refine and test our method. Currently our DL method is approaching expert human accuracy for classifying patients into those with and those without imaging features indicative of ischaemic stroke. Our method development has included approaches to better understand how the DL system makes a classification of ischaemic stroke vs no ischaemic stroke and to enable the system to highlight areas of interest on the images. Our method is novel for several reasons:

  1. Our comprehensive pre-processing data pipeline is designed to handle a range of patient-related and technical imaging variabilities common in routine clinical practice (i.e. it is expected to be robust for handling real CT data that are frequently imperfect).
  2. We used non-annotated but labelled images for DL development. In other words, we did not draw round the stroke lesions but let the system identify relevant features independently after we provided scan labels, separate from the actual images, to indicate to the system which scans were positive or negative for features of stroke, based on expert interpretation.
  3. Our system assesses the whole brain for features of ischaemia and, in contrast to most commercially available tools, does not just focus on the most commonly affected areas such as the territory of the middle cerebral artery.

Next steps are to incorporate common background brain features (such as reduced brain volume, old strokes, and other measures of chronic brain injury) into the DL model. This may help to improve the accuracy of the model, but will also allow us to explore relationships between CT brain imaging features identified by our system at baseline and future health of the patient.

Ultimately, we hope to provide a system that can work robustly and intelligently in a routine clinical environment to identify acute and chronic brain changes relevant to stroke, that might be used at the individual patient level to guide decisions about treatment.

Project Milestones


  • Converted and anonymised over 5000 CT image sets from DICOM (the standard format for all medical imaging) to a format more suitable for DL.
  • Securely transferred anonymised data to a university network with access to GPU and high performance processing capabilities.
  • Created a data pipeline for pre-processing CT brain imaging (to standardise, quality assess and minimise data loss of image sets used for DL development and testing).
  • Presented details of substantial data preparatory steps at national and international AI and stroke meetings, respectively.
  • Developed and refined a neural network to learn features of ischaemic stroke from the pre-processed and labelled CT image sets.
  • Applied methods to enable interpretation of DL outputs and to highlight areas of interest on individual CT scans.
  • Compared DL system results against a panel of expert human interpretations of the same scans.

Not yet completed:

  • At least 2 publications so far – Current drafts in development:
    1. Detailing data preparatory steps and CT processing pipeline
    2. Development and testing of DL method for ischaemic stroke feature detection, comparison with experts, in relation to other available methods.
  • Incorporation of expert labels for background brain changes.

Development and testing of secondary methods to predict outcome after stroke using only data available at baseline

Project Team / Collaborators

  • Joanna Wardlaw & Grant Mair (Centre for Clinical Brain Sciences, University of Edinburgh)
  • Amos Storkey, Wenwen Lee, Alessandro Fontanella, Antreas Antoniou, Eleanor Platt & Elliot Crowley (School of Informatics, University of Edinburgh)
  • Emanuel Trucco (School of Science and Engineering, University of Dundee)