Optimisation and clinical validation of a digital diagnostic system for small bowel biopsies

Question: This proposal addresses a significant unmet need, for accurate and timely duodenal biopsy diagnosis.

Background: The international shortage of histopathologists and lack of automation in histopathology cause delays in biopsy diagnosis and backlogs, with knock-on effects for patient care.

There is disagreement between histopathologists about the diagnosis of 20% of duodenal biopsies, particularly in diagnosing coeliac disease.

While 1% of the population has this diagnosis, 2% remains undiagnosed, with symptoms including abdominal pain, diarrhoea, vomiting, malaise, fatigue, mouth ulcers and itchy skin rashes.

Longer term complications include duodenal adenocarcinoma, lymphoma, vitamin deficiency, anaemia, osteoporosis and infertility.

Innovate UK-funded infrastructure introduced high throughput scanning of microscope slides, which histopathologists can view on a screen, rather than under a microscope. The scanned images provide the opportunity to automate biopsy diagnosis.

Lyzeum has developed software, with >96% accuracy, that uses multiple instance learning, breaking the large scanned images into small “tiles”, and predicts likely diagnosis of each tile, to classify biopsy images based on the percentages of tiles with particular diagnostic labels.

Aims: 1.We will optimise and validate our software solution for the fully automated reading of duodenal biopsies, dividing them into &ldquo;normal&rdquo; (&asymp;73%), coeliac disease (&asymp;12%) and &ldquo;for histopathologist review&rdquo; (<30%). 2.We will determine performance metrics: percentage of biopsies amenable to fully automated diagnosis, sensitivity, specificity, accuracy. 3.Immediately after the study, we will embed the software, in a commercial image-handling platform, permitting a small regulatory study for MHRA approval, ready for clinical adoption.

Methods: Using 3 NHS Histopathology Departments that utilise different scanning and image-handling systems (to maximise generalisability), we will add 7200 new clinicopathologically annotated images to our existing 4000-image dataset, for software training. Using cross-validation, we will identify images that persistently misclassify.

Our histopathologists will review these images and associated clinicopathological data, to maximise ground truth accuracy. We will collect 800 additional images for testing, to generate performance metrics. Timelines: After a 2-month setup phase, we will collect 12 months image data. A 4-month post-collection biopsy review phase will maximise ground truth accuracy.

Software training will occur in months 2 &ndash; 21, testing in months 22-23 and result collation in month 24, in preparation for commercial engagement.

Anticipated impact: We anticipate a wide range of impacts on health outcome and health economics, benefitting patients, the NHS and public health.

Eradicating pathologist-to-pathologist health inequality in duodenal biopsy diagnosis, we envisage improving diagnostic accuracy from 80% (histopathologists) to 96% (automated).

This will avoid false negative/equivocal results and chronic, debilitating symptoms and complications in patients with coeliac disease.

Eradication of delays in biopsy diagnosis will avoid delayed decisions about patient management and facilitate efficient use of NHS resources.

Offloading some of pathologists workloads, particularly in severely understaffed departments, will aid recruitment and retention. Successful NHS adoption of our solution will provide a stepping-stone to broader automation in Histopathology.

Dissemination: We will publish in scientific journals, present data at conferences (Pathological Society of Great Britain and Ireland, British Society for Gastroenterology, Coeliac UK, Global Engage and DigiBase) and disseminate information to patient groups.