Automating the diagnosis of small bowel biopsies: plugging the gap of gluten sensitivity

Question: To improve the accuracy of duodenal biopsy diagnosis, can we improve the performance of our software in detecting coeliac disease in patients on a gluten-free diet? Background: Coeliac disease (CD) is typically diagnosed by a pathologist examining a duodenal biopsy under the microscope. To ensure characteristic changes of CD in gluten-sensitive patients’ biopsies, patients must consume a gluten-rich diet for 6 weeks prior to endoscopy. This often results in debilitating symptoms and poor compliance. Insufficient gluten consumption leads to biopsies looking normal and CD diagnoses being missed. Undiagnosed CD causes unpleasant gastrointestinal and systemic symptoms, with a risk of serious complications (duodenal lymphoma and carcinoma, osteoporosis and infertility). Duodenal biopsies from CD patients re-biopsied after 6 months on a gluten-free diet (denoted CD-GFD) act as a proxy for patients who consumed insufficient gluten. Beginning preparations for regulatory approval (Quality Management System and framework for regulatory application) will help us obtain future translational research grant or venture capital funding. Aims: 1. To improve the performance of our software, by training it on CD-GFD biopsies 2. To make our software available to pathologists on a research-use-only basis 3. To develop a Quality Management System 4. To develop a framework for regulatory approval of using AI software in diagnostic pathology for the automated diagnosis of “normal” Methods: 1. Data collection (months 1-5): we will collect 800 duodenal biopsy images (20% GFD-CD, 20% active CD (positive control), 40% normal (negative controls) and 20% other abnormality (neither CD nor normal)) from 4 NHS centres (Cambridge (200), Southampton (100), Edinburgh (100) and Oxford (100)) and 1 pathology image consortium (NPIC, 300 biopsy images). 2. Software development, testing and evaluation (months 1-10): we will train our software on CD-GFD images to maximise classification accuracy, undertaking cross-validation, and then testing it on a holdout set (20% biopsies). 3. ?5 volunteer pathologists will also diagnose the holdout set, and then re-diagnose it after receiving the software’s diagnosis and measurements for key parameters, to determine how the software works as a pathologist’s assistant. 4. Pathologist feedback (5-12): we will integrate our software into a widely used pathology imaging platform (Sectra), for pathologists to trial it (for “research use only”) and provide valuable user feedback. 5. Regulatory approval (months 1-9): we will develop a Quality Management System. We will also identify automated diagnostic approaches in clinical laboratories and radiology, to develop our approach for a regulatory application for automating diagnosis in histopathology. Impact: AI can reduce the current 20-25% inter-observer disagreement in CD diagnosis, avoiding missed diagnoses, reducing long-term morbidity. Future clinical use of our AI system automating the diagnosis of most normal and CD duodenal biopsies could free up >16,000 hours of UK pathologists’ time, which can be spent clearing backlogs of other specimen types. Dissemination: We will publish our software evaluation, including the comparison with 5 experienced pathologists (with and without AI assistance) in a high-impact medical journal. We will present work at medical conferences. With Coeliac UK, we will run a PPIE workshop, and make a video and write an article for patients.