Deep learning for assessing brain imaging in stroke

  • Research type

    Research Study

  • Full title

    Automated detection and classification of imaging features representative of stroke on brain CT using computational deep learning

  • IRAS ID

    257546

  • Contact name

    Grant Mair

  • Contact email

    grant.mair@ed.ac.uk

  • Sponsor organisation

    ACCORD

  • Duration of Study in the UK

    0 years, 6 months, 0 days

  • Research summary

    For patients presenting to hospital with symptoms of stroke, a non-enhanced CT scan of the brain is performed rapidly to determine the likely underlying pathology and assess patient suitability for treatment. Treatments vary considerably for patients with ischaemic versus haemorrhagic stroke versus other pathologies that can mimic stroke (e.g. brain tumours) and could be harmful if administered incorrectly. The initial assessment of this brain CT may fall to clinicians with limited radiological experience. Expert radiological interpretation is not routinely available at all hours.

    Ischaemia is the commonest cause of stroke symptoms, affecting up to 85% of patients. Eligibility for treatment with intravenous alteplase or mechanical thrombectomy is largely time dependant. For the majority of patients, treatment with alteplase is restricted to 4.5 hours from stroke onset in Europe while thrombectomy is routinely offered <6 hours from onset.

    We propose to develop a computational deep learning system for the automated assessment of standard brain CT, the most common imaging modality acquired for patients with symptoms of stroke. By first recognising and differentiating the three major types of underlying pathology that present with symptoms of stroke (ischaemia, haemorrhage, brain tumour), this initial iteration of our system will be unique and will act as the foundation upon which to build other bespoke features designed to support clinicians make appropriate and safe treatment decisions using CT.

    Deep learning is a form of machine learning where computer algorithms use artificial intelligence to learn patterns from data rather than relying on explicit human programming. Convolutional neural networks (CNN) are at the forefront of deep learning as applied to medical imaging datasets. We will use state-of-the-art CNN architectures to learn from large (total ~3500 patients), validated and expert labelled CT imaging datasets representative of routine clinical practice in stroke.

  • REC name

    London - Queen Square Research Ethics Committee

  • REC reference

    18/LO/2175

  • Date of REC Opinion

    6 Dec 2018

  • REC opinion

    Favourable Opinion