The challenge is to develop an AI or machine-based learning programme that can help healthcare organisations accurately identify whether a patient has a fracture. This is initially a classification problem (by assigning a value of yes, no or maybe).
In simple terms, the task to begin with is to develop an automatic system that, with a degree of certainty, can remove from clinicians’ workload those that are definitely yes or no, leaving them to focus on the more complex images. This is an initial step towards integrating AI systems into a mainstream clinical workflow within the NHS and could be a platform for building more intelligent learning systems.
Each year in Scotland, the NHS gives some 5,000 patients x-rays of the peripheral upper limb (wrist or hand) and lower limb (ankle or foot), 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 morbidity and financial cost.
The interpretation of peripheral limb x-rays is the remit of a wide variety of clinical staff in many clinical settings, from large urban emergency departments to nurse-led remote cottage hospitals and minor injury units.
The diagnosis of a fracture in the wrist or ankle is made from 2 standard radiographic views taken at right angles to each other. Radiographic fracture assessment of the hand or foot may include a third oblique view.
Recently published studies have successfully used machine learning to analyse radiographs to detect fractures. They have shown the ability to perform at the same diagnostic standard as an expert.
AI or machine learning could be included in clinical workflows to interpret peripheral limb radiographs for the presence of fractures, which in most cases are not reported for several days. This would help:
- improve diagnostic accuracy and treatment
- improve patient pathways and outcomes
- reduce the growing deficit between radiology reporting workloads and staffing levels
This competition draws on Scotland’s expertise in:
Example: in 2019
- clinical and academic digital radiology
- advanced data storage
- data governance and access
- interoperable healthcare databases
John Doe has a swollen right wrist after falling on an outstretched hand in the street. He lives in a rural location and attends his local minor injury unit where he is seen by a nurse who requests x-rays.
The films are placed on a digital archiving system but both the nurse and the radiographer are unsure if there is a fracture. Mr Doe is keen to get back to his activities, including driving. The staff decide to let him keep his wrist free but say they will contact him if an abnormality is found on the formal radiology report. After 4 weeks the formal radiology report shows there was a fracture. Mr Doe is recalled and given more x-rays. They show the bones have moved and he will need an operation.
The initial reading of the radiology image is made by a variety of grades of staff with wide experience. While local departments often have safeguard systems to minimise risk, some fractures have a delayed diagnosis. Peripheral limb injuries may have significant morbidity and often financial or lifestyle implications.
Example: in 2020
Following the application of machine learning, John Doe’s x-ray is displayed to the clinician with an augmented image highlighting the presence of a fracture or abnormality. The clinician can use this information alongside other clinical details and, if necessary, seek a specialist review.
If no fracture is found, patient management is simpler as the films do not have to go for formal radiology reporting.
This real-time system of augmentation has:
- significantly and reliably improved the confidence the patient and the clinician have in the diagnosis of ‘no fracture’
- reduced the number of specialist consultations for patients with a suspected fracture who do not have a fracture
- sizeably and safely reduced radiology reporting, letting the department concentrate on more complex image interpretation
Successful applicants must use an available dataset of peripheral limb x-rays and linked text-based reports from the University of Aberdeen’s accredited secure Grampian Data Safe Haven (DaSH). With these they will develop AI algorithms to:
- interpret the existing text-based report to categorise as fracture or no fracture
- interpret the radiograph image to identify the presence of fracture
- develop an AI product with the required level of real world accuracy to enhance to enhance radiology image interpretation in mainstream clinical practice
The competition is looking for proposals that:
- improve peripheral limb fracture detection by non-radiology experts in out of hours environments within NHS Grampian
- transform peripheral limb injury clinical pathways to improve patient outcomes and increase productivity by at least 20%
- use the relevant NHS, academic and commercial expertise, data and infrastructure offered by Grampian
- have clinical and commercial potential locally, nationally and globally
We are looking for industrial innovators. You must confidently collaborate and use multiple data sources to develop clinically relevant and commercially practicable solutions. There is potential to commercialise outputs directly through NHS Scotland and globally through the sales and marketing channels of Canon Medical.
Any adoption and implementation of a solution from this SBRI competition would be the subject of a separate, possible competitive, procurement exercise. This competition only covers R&D not the purchase of any solution.
The performance of models trained on the dataset that is made available during the programme will be validated against an unseen dataset. There will be a further dataset available to demonstrate generalisability.