CARNAX
AI for detection of intestinal perforation.
Modality:
Abdominal and Chest X-Rays (Paediatric)
Pathology:
Perforation of the intestine in preterm neonates
Status:
Concluded
CSC Lead: Mike
This project was extensively worked on for several years but in the end had to be withdrawn. Primary data sources for this project were mobile Xrays which varied in contrast and quality, size and positioning. The term ‘pre-term neonates’ includes babies of significant differences in size and maturity. Due to the heterogeneity of data, sufficient AI performance could not be achieved. This project could potentially be revived if medical AI technology in computer vision advances, but for now this project is concluded.
About one in every ten newborn babies requires support on a neonatal unit. Most of these infants are premature (born at less than 37 weeks of pregnancy). Premature infants are at greater risk of intestinal problems which can lead to infection, inflammation or a hole in the wall of the bowel. Clinical teams that look after these babies rely on x-rays of the abdomen to know if there is a problem with a child’s gut. On occasion problems are missed because the x-ray is not looked at carefully enough or because the clinician is tired or inexperienced. A computer algorithm can be trained to recognise abnormalities on an x-ray. This system can then be used to alert the clinical team early if there is a problem. This could prompt earlier treatment and prevent babies from getting sicker.
Clinical lead: Hammad Khan
Rationale
Bowel perforation in neonates is difficult to see on Xrays. Because of radiosensitivity and frequency of other complications, the timely diagnosis and intervention are crucial. GSTT is a specialist centre where many of these babies have been seen and thus we sit in the unique position of being able to develop a lifesaving tool like this.
Patient pathway
Preterm newborn babies with suspected bowel perforation are referred for an abdominal Xray scan. The Xray is acquired on a mobile Xray machine, with dose optimised for radiation safety of the neonate. The images are reported by dedicated consultant radiologist. Where a consultant is not available (e.g. at night), and the on-call registrar suspects a perforation, a consultant is called. If the bowel perforation is not demonstrated, the registrar makes the report.
Training data
Approximately 2000 Xrays were included from both GSTT and KCH.
Risks
Low performing AI could give false confidence in a wrong result. Biggest risk is in the over-reliance on AI.
Goals
Increase confidence in reporting of on-call non-experts.
Success criteria
AI tool which demonstrates high sensitivity and high specificity.
Alternatives
Currently no commercial products identified.