The Imperial AI Echocardiography Dataset

  • Research type

    Research Database

  • IRAS ID

    279328

  • Contact name

    Matthew Shun-Shin

  • Contact email

    m.shun-shin@nhs.net

  • Research summary

    The Imperial AI Echocardiography Dataset

  • REC name

    South Central - Oxford C Research Ethics Committee

  • REC reference

    20/SC/0386

  • Date of REC Opinion

    2 Dec 2020

  • REC opinion

    Further Information Favourable Opinion

  • Data collection arrangements

    This project will create an open-access database of fully-anonymised ultrasound images of the heart to further the development of AI for detecting heart disease.

    Ultrasound images of the heart (echocardiograms) will be retrieved from the hopital datastore and undergo anonymisation, removing any image and file meta-data the identifes a patients (such as name, date of birth, hospital number) and collated.

    These images will then have the cardiac structures (such as the muscle or valves) labelled by experts.

    The anonymised images and labels will then form part of a dataset to be make openly-available to advance the developmen of machine learning and AI.

  • Research programme

    Echocardiography provides a huge amount of clinical information that can directly and immediately influence patient care. Machine learning algorithms can support the use of echocardiography by general and acute physicians, aiding with image acquisition and interpretation so they can rapidly diagnose severe cardiac disease. Some of the diseases which artificial intelligence (AI) can be trained to recognise are thickened valves which struggle to open (stenosis), leaking valves (regurgitation), weak hearts which are not pumping properly (systolic dysfunction), increased wall thickness (hypertrophy), and fluid around the heart (pericardial effusion). Neural networks are trained by showing them large numbers of images and their associated labels). Within the machine learning community, there are lots of large open-access datasets of labelled images (e.g. ImageNet has 14 million images with labels). However, there are very few open-access datasets of labelled medical images, and in particular cardiac imaging. We will develop a freely available a fully anonymised and annotated dataset of echocardiography images to stimulate further research in this area.

  • Research database title

    The Imperial AI Echocardiography Dataset

  • Establishment organisation

    Imperial College Healthcare NHS Trust

  • Establishment organisation address

    Hammersmith Hospital

    Du Cane Road

    London

    W12 0HS