@campathology.bsky.social
@hugheshall.bsky.social
@uniofcam.bsky.social
@cuh.nhs.uk
@ai.cam.ac.uk

A big thank you to all of my amazing collaborators and all of our incredibly partners and funders who made this work possible
@coeliacuk.bsky.social
@innovateuk.bsky.social
@acceleratescience.bsky.social
@cambridgec2d3.bsky.social
@cambridgebrc.bsky.social
@nihr.bsky.social

We hope that our interpretable AI approach marks a significant first step toward software that could support faster, more accurate and consistent coeliac disease diagnosis. [5/5]

Evaluating the IEL-to-enterocyte (in the villi and the crypts) and villus-to-crypt ratios on a large independent test set from a previously unseen hospital, we observed statistically significant differences for all ratios between the normal and coeliac disease populations. [4/5]

We showed how the models can accurately predict the IEL-to-enterocyte ratio. An increased IEL-to-enterocyte ratio is a key indicator of coeliac disease. However, unlike pathologists who only have time to count a few cells, the AI model can detect 1000s of cells in the entire biopsy in seconds. [3/5]

We developed segmentation models that can identify villi, crypts, intraepithelial lymphocytes (IELs), and enterocytes in H&E-stained duodenal biopsies, the four key structures used by pathologist when diagnosing coeliac disease. [2/5]

@campathology.bsky.social, @hugheshall.bsky.social, @uniofcam.bsky.social, @nejm.org, @ai.nejm.org, @cuhnhs.bsky.social, @ai.cam.ac.uk

A big thank you to all of my amazing collaborators and all of our incredibly partners and funders who made this work possible
@coeliacuk.bsky.social,
@innovateuk.bsky.social,
@acceleratescience.bsky.social,
@cambridgec2d3.bsky.social,
@cambridgebrc.bsky.social

If you would like to read more please see our paper (ai.nejm.org/stoken/defau...), or if you prefer a less technical article see this Guardian article (www.theguardian.com/science/2025...), or this report written by the University Comms team (www.cam.ac.uk/stories/AI-a...)

We compared diagnoses from four experienced pathologists with those from our AI model. The result: average agreement between any two pathologists was identical (90%) to the agreement between each pathologist and the AI model.
The takeaway: the AI matches expert-level accuracy.

Our model was trained and validated on 3,000+ cases from 4 hospitals using 5 different scanners.
When tested on 600+ cases from a completely unseen hospital, it maintained high accuracy across all adult patient subgroups, demonstrating strong generalisability in real-world settings.

Our model pipeline looks as follows:
i) remove artefacts + separate tissue from background
ii) break down biopsy image into 256x256 pixel sub-patches
iii) apply stain normalisation
iv) train a ResNet model with multiple-instance learning
v) run inference + generate heatmaps to visualise prediction

@campathology.bsky.social @hugheshall.bsky.social @uniofcam.bsky.social @nejm.org @ai.nejm.org @cuhnhs.bsky.social @ai.cam.ac.uk

A big thank you to all of my amazing collaborators and all of our incredibly partners and funders who made this work possible @coeliacuk.bsky.social @innovateuk.bsky.social @acceleratescience.bsky.social @cambridgec2d3.bsky.social @cambridgebrc.bsky.social

If you would like to read more please see our paper (ai.nejm.org/stoken/defau...), or if you prefer a less technical article see this Guardian article (www.theguardian.com/science/2025...), or this report written by the brilliant University Comms team (www.cam.ac.uk/stories/AI-a...)

We compared diagnoses from four experienced pathologists with those from our AI model. The result: average agreement between any two pathologists was identical (90%) to the agreement between each pathologist and the AI model.
The takeaway: the AI matches expert-level accuracy.

Our model was trained and validated on 3,000+ cases from 4 hospitals using 5 different scanners.
When tested on 600+ cases from a completely unseen hospital, it maintained high accuracy across all adult patient subgroups, demonstrating strong generalisability in real-world settings.

Our model pipeline looks as follows:
i) remove artefacts + separate tissue from background
ii) break down biopsy image into 256x256 pixel sub-patches
iii) apply stain normalisation
iv) train a ResNet model with multiple-instance learning
v) run inference + generate heatmaps to visualise prediction