AI-supported early fracture diagnosis (Phase 2: Bering Limited)

• Research type

  Research Study

• Full title

  Artificial intelligence-supported early fracture diagnosis (Phase 2: Bering Limited)

• IRAS ID

  288197

• Contact name

  Ignat Drozdov

• Contact email

  idrozdov@beringresearch.com

• Sponsor organisation

  Bering Limited

• Clinicaltrials.gov Identifier

  DaSH 393, DaSH SafeHaven Number

• Duration of Study in the UK

  1 years, 0 months, 0 days

• Research summary

  Summary of Research

  Each year in Scotland, the NHS gives some 5,000 patients x-rays of wrists, hands, ankles and feet, most often looking for a fracture after trauma. Although isolated injuries in these areas are often categorised as ‘minor’, misdiagnosis and consequent mismanagement can result in significant impact for patients and financial costs to the NHS.
  Artificial Intelligence (AI) or “machine learning” (a set of procedure rules to take in clinical data such as X-rays, assess the risk of a fracture and present this risk and information to a clinical team) could be developed to help clinicians make diagnoses.
  To develop AI or machine learning tools and to take these tools to the level of “approved for health care use” and integrated into the appropriate IT and/or equipment for healthcare use requires a partnership between NHS, academia and industry.
  NHS Grampian A&E and Radiology clinicians have identified that there is significant clinical need and are eager to work in partnership with those with the technical skills to develop potential solutions.

  A pilot study (Phase 1) involving a small (100 patients) fully anonymised x-ray dataset (no patient names, addresses, date of birth or hospital numbers) has already been completed (IRAS no. 271600) using an AI laboratory space within the Grampian Data Safe Haven (DaSH). After successfully demonstrating potential to develop useful clinical solutions to support fracture detection, we have been invited to proceed to Phase 2.
  In a second phase, we will access a larger dataset (images for 10,000 patients) to develop our solutions further. We will create a more robust natural language processing model for annotation of free-text radiological reports and we will scale our deep convolutional neural network architectures to detect and localise fractures in peripheral limb x-rays.

  Summary of Results
  The aim of the project was to deliver an automated Artificial Intelligence system for fracture detection on peripheral limb X-Rays.

  The developed AI system, trained on nearly 50,000 X-Rays, takes a single view acquired from a patient ≥ 16 years of age as a DICOM file, and produces several outputs:
  • Quality Control (QC) assessment of the input image,
  • An output score between 0 and 100 related to the likelihood that the X-ray is radiologically normal
  • Output scores between 0 and 100 for either Acute Fracture or non-Fracture abnormality (e.g. osteoporosis)
  • A heatmap indicating areas of likely abnormalities

  The model achieves Area Under Precision Recall curve of 0.94 for Fracture detection, reporting 80% of normal X-Rays with 100% accuracy. A clinical use of our AI model on a retrospectively-collected dataset of n=200 wrist and ankle X-Rays by two Consultants ED clinician delivered 32.2% and 16.4% improvements to baseline clinician sensitivity and specificity, compared to Consultant Radiologist. These findings suggest that:

  1) Our AI model can be used to auto-report normal peripheral limb X-Rays, delivering operational benefits in busy emergency departments,
  2) Concomitant usage of our AI predictions with clinical judgement augments clinician performance, and
  3) Our model achieved AU PR of 0.92 on a collection of n=1000 X-Rays from NHS GGC, suggesting that the results are generalisable.

  Overall, there have been no variation in planned and delivered outputs

• REC name

  East of Scotland Research Ethics Service REC 2

• REC reference

  20/ES/0085

• Date of REC Opinion

  20 Aug 2020

• REC opinion

  Favourable Opinion