Small Bowel Deep Learning Algorithm Project
- Conditions
- Crohn Disease
- Interventions
- Other: Machine learning algorithm
- Registration Number
- NCT03706664
- Lead Sponsor
- London North West Healthcare NHS Trust
- Brief Summary
Crohn's disease affects 200,000 people in the UK (\~1 in 500), most are young (diagnosed \< 35 years) with costs of direct medical care exceeding £500 million.
Crohn's disease is caused by an auto-immune response and affects any part of the digestive tract, most commonly the last segment of the small bowel (the terminal ileum).
Magnetic resonance imaging (MRI) plays a role in 3 areas: Crohn's disease diagnosis , monitoring treatment response \& assessing development of complications.
To evaluate the small bowel using MRI, Radiologists visually examine the scan slice-by-slice. The interpretation is time consuming and error-prone because of disease presentation variability and differentiation of diseased segments from collapsed segments.
Deep learning for image analysis is based on a computer algorithm "learning" from human (Radiologist) generated training data.
This method has been successfully applied to medical imaging, for example computer detection of lung cancer on chest X-rays.
This pilot study investigates if a deep learning algorithm can identify and score segments of inflamed terminal ileum affected by Crohn's disease.
To our knowledge this is the first project attempting to develop such an algorithm.The study will retrospectively review MR images obtained as part of standard care from patients being investigated for, Crohn's or being followed up with Crohn's disease. 226 patients' images will be used for the study.
On fully anonymised images two Radiologists working at Northwick Park Hospital will score and outline normal and abnormal loops of terminal ileum. Imperial College computer science department will then develop a deep learning algorithm from imaging features of normal and abnormal loops.
The study end-point is algorithm performance vs. images labelled by Radiologists.
The eventual aim is to develop an algorithm that assists Radiologists in the accurate diagnosis and follow-up of patients with Crohn's disease.
- Detailed Description
Introduction.
The principal aim of the study is evaluating the accuracy of deep learning algorithm in differentiating between normal and abnormal terminal ileum against experienced Radiologists on MR Enterography images.
The study builds on existing research, which has shown statistical methods can identify sites of small bowel Crohn's disease. However the process was time consuming \>1hr and not fully automatic. Our project investigates if cutting edge "deep learning" algorithm (based on neural networks) coupled with increased computing power can provide accurate and timely information.
The project has been designed jointly by Specialist Radiologists in Gastrointestinal imaging (who are aware of the challenges in imaging Crohn's disease accurately) and Imperial College Computer Science Department (who are experienced in developing neural networks for medical imaging). Input and review from London North-West Research and Development department is also acknowledged.
Study design.
Retrospective design \& Recruitment.
The study will retrospectively identify eligible patients and use a consecutive case sampling technique, (all eligible images will be included working backwards from most recent).
This retrospective approach compromises between generalisability of findings being reduced vs. the study being carried out relatively quickly and at low cost (study has no grant funding).
The investigators are confident of the generalisability of the results as a recruitment target of 113 normal cases and 113 cases with terminal ileal disease should cover the spectrum of normal and abnormal appearances (previous studies have used \<50 image sets).
Cases with normal terminal ileum on MRI are included as an approach to algorithm development involves comparison of imaging features of normal and abnormal terminal ileum on MRI studies.
Non-experimental approach.
The method uses MRI scans undertaken as part of standard clinical care. No additional imaging is undertaken for this study. The study results will not change the current treatment/s eligible patients are taking..
Consent \& confidentiality.
As the images used for algorithm development are fully anonymized so explicit consent will not be obtained. This follows guidance from the General Medical Council Guidelines in 2011 and The Royal College of Radiologists(UK) in 2017 which state anonymized recordings can be shared for use in research without consent.
MRI images used for this study were acquired as part of routine standard clinical care, and would routinely be viewed by the Radiologists taking part in this study as part of their normal working practice.
As soon as suitable patient is identified the patient's images will be copied in a fully anonymized form with no direct or indirect identifiers. A robust anonymization function is included in the Radiology image viewing program. Study subject Identifiers will be randomly allocated preventing pseudo-anonymization if scans from the same patient at different time points are included.
No sensitive/patient identifiable data will be transferred for algorithm development during the study. The algorithm development is based on matching MRI pixel intensities to disease scores/ annotations across multiple scans. Anonymization does not affect the pixels within the image. Only aggregate data will be presented in publications- i.e. single case examples will not be published.
Conflict of interest. The researchers on this study declare no conflict of interest.
Recruitment & Eligibility
- Status
- ACTIVE_NOT_RECRUITING
- Sex
- All
- Target Recruitment
- 226
Not provided
Not provided
Study & Design
- Study Type
- INTERVENTIONAL
- Study Design
- SINGLE_GROUP
- Arm && Interventions
Group Intervention Description Training of machine learning algorithm Machine learning algorithm 113 MR Enterography images labelled by Radiologists will be used to develop a machine learning algorithm to (1) localise the terminal ileum, (2) classify the terminal ileum as normal or abnormal. Testing of machine learning algorithm Machine learning algorithm 113 MR Enterography images labelled by Radiologists will be used to test the accuracy of the machine learning algorithm to (1) localise the terminal ileum, (2) classify the terminal ileum as normal or abnormal compared to Radiologists opinion. Cross Validation analysis will be used for data analysis.
- Primary Outcome Measures
Name Time Method Machine learning algorithm's ability to accurately localize the terminal ileum. 24 months Study will compare manually segmented regions of interest by Radiologists with predictions by machine learning localisation algorithm.
- Secondary Outcome Measures
Name Time Method Data processing time until a diagnosis reported by algorithm. 24 months Study will assess time taken for algorithm to give a diagnostic outcome. (Previous studies have shown this time can be variable).
Machine learning algorithm's ability to accurately distinguish abnormal and normal terminal ileum. 24 months Agreement between Radiologists and predictions by machine learning classification algorithm will be analysed.
Trial Locations
- Locations (1)
St Mark's Hospital
🇬🇧London, Harrow, United Kingdom