MIDI (MR Imaging abnormality Deep learning Identification)

  • Research type

    Research Study

  • Full title

    Deep Learning for Identification of Abnormalities on Head MRI

  • IRAS ID

    235658

  • Contact name

    Thomas Booth

  • Contact email

    thomasbooth@nhs.net

  • Sponsor organisation

    King's College London

  • Clinicaltrials.gov Identifier

    NA, NA

  • Duration of Study in the UK

    5 years, 0 months, 1 days

  • Research summary

    There is a wide variation in how incidental findings (IFs) discovered in ‘healthy volunteers’ are managed. Routine reporting of ‘healthy volunteer’ scans by a radiologist is a challenging logistic and financial burden. It would be valuable to devise automated strategies to ensure that IFs can be reliably and accurately identified potentially removing 90% of scans requiring routine radiological review, thereby increasing the feasibility of implementing a routine reporting strategy.

    An automated strategy could also address the unmet clinical need in identifying abnormalities quicker, potentially allowing for early intervention to improve short and long-term clinical outcomes. Radiologist shortages and increased demand for MRI scans mean delays in reporting, particularly in the outpatient setting.

    Deep learning is a new technique in computer science that automatically learns hierarchies of relevant features directly from the raw inputs (such as MRI or CT) using multi-layered neural networks. A deep learning algorithm will be trained on a large database of head MRI scans to recognise scans with abnormalities. This algorithm will be trained to classify a subset of these scans as normal or abnormal. The technique will then be tested on an independent subset to determine its validity.

    If the tested neural network has a high diagnostic accuracy, future research participants may benefit as currently not all institutions review their research scans for incidental findings. Similarly, in those cases where scans may not be reported for weeks, patients may benefit. In both research and clinical scenarios, an algorithm would quickly identify abnormal pathology and prioritise scans for reporting.

    In summary, the aim is to develop a deep learning abnormality detection algorithm for use in both the research and clinical setting.

  • REC name

    Yorkshire & The Humber - Sheffield Research Ethics Committee

  • REC reference

    18/YH/0458

  • Date of REC Opinion

    3 Dec 2018

  • REC opinion

    Further Information Favourable Opinion