UCL Great Ormond Street Institute of Child Health


Great Ormond Street Institute of Child Health


Artificial Intelligence for Paediatric Fracture Detection

Supervisors: Owen Arthurs, Susan Shelmerdine, Baris Kanber

Review of the Key Literature:
Fractures are extremely common, with up to a half of all children sustaining a fracture at some point during childhood (~133.1 per 10,000 p.a). They are a leading cause for long-term disability and present in 55% of children who have been physically abused. Unfortunately detecting fractures on children’s radiographs can be challenging because they are subtle, there are a wide range of normal appearances and injury patterns differ to adults.

Hypothesis and Aims:
The aim of this study is to develop a deep learning model capable of fracture detection across multiple body parts to the same accuracy as an expert paediatric radiologist and to determine whether this would improve clinician decision making in practice.

6-12 month plan:
In the first 12 months of the studentship, anonymised Xray imaging data will be collected and labelled by radiologists from 8 different UK centres. Some of the Xray images will be abnormal, some normal. Using a pre-trained deep learning model, we will enhance and adapt this to identify and locate abnormal bones on children’s X-rays and use them to predict abnormalities on previously unseen data. Analysis will be conducted based on different body parts on Xray imaging and include at-risk groups (e.g cases with brittle bone disease and suspected abuse).

After the first 12 months, we will proceed to try to find out if this imaging helps improve clinicians performance when using AI to aid their usual diagnosis in a clinical simulation study. There will be opportunities to work with industry and AI collaborating partners as part of this project.

By year 1: Secure well curated library of multi-centric anonymised limb radiographs
By year 2: Preliminary fracture detection model for paediatric imaging
External validation benchmarked against expert radiologist readers.

By year 3: Simulated clinical implementation study & key findings presentation at international and national conferences.

Ethics Approval:
Ethical approval has been granted for ‘Use of routine data for research’ (IRAS ID:214972) and ‘Non intervention evaluation of digital technologies to improve outcomes in children attending hospital’ (IRAS ID:235646).

HRA approval for anonymised multicentric data collection from paediatric trauma centres has been granted and the R&D administrative process at these centres has recently begun.

Recommended Reading/References:
1.  Kuo RYL et al. Artificial intelligence in fracture detection: a systematic review and meta-analysis. Radiology. 2022 Mar 29; 211785.
2.  Cohen JF, McInnes MDF. Deep learning algorithms to detect fractures: systematic review shows promising results but many limitations. Radiology 2022 Mar 29; 212966.
3.  Dupuis M et al. External validation of a commercially available deep learning algorithm for fracture detection in children. Diagn Interv Imaging. 2022 Mar; 103(3):151-159.
4.  Shelmerdine SC et al. Artificial intelligence in paediatric radiology: international survey of health care professionals’ opinion. Pediatr Radiol. 2022 Jan; 52(1):30-41.
5.  Choi JW et al. Using a dual-input convolutional neural network for automated detetion of pediatric supracondylar fracture on conventional radiography. Invest Radiol. 2020 Feb; 55(2):101-110