Can AI score Crohn disease from a capsule video?
A Shanghai model combined lesion detection with sequence analysis and tracked expert Lewis scores in small internal tests. It sampled only about 7% of a typical video, used one hospital and one capsule system, and has not been tested in a prospective clinical workflow.
The short version
3 key points
- 1.The lesion component used 725 patient cases and 201,860 retained images; the Lewis-score component used only 89 cases, split into 49 for training and 40 for internal testing.
- 2.Internal score testing produced mean absolute error 2.9 and R² 0.82. A 20-case later cohort from the same hospital produced mean absolute error 3.3; its near-perfect R² is not external validation.
- 3.The sequence model analysed 900 frames—about 7.1% of a typical 12,600-frame small-bowel recording at the study's sampling rate—and has not shown that it saves clinician time or changes treatment safely.
The answer: the system reproduced expert scores in one centre, but it is not ready to score patients on its own
Eval-Model is an unusually complete attempt to move beyond finding isolated lesions in capsule-endoscopy images. It detects and localises ulcers and oedema, classifies stenosis and passes those features through a sequence model to estimate the examination-level Lewis score used to describe small-bowel inflammatory activity in Crohn disease. In a 40-case internal test, the predicted scores had a mean absolute error of 2.9 and an R-squared of 0.82 against expert consensus.
The result supports technical feasibility at the Children's Hospital of Fudan University. It does not establish clinical superiority, safety or transportability. The scoring model learned from 49 examinations, was tested on 40, and was then checked on only 20 later examinations from the same hospital, the same capsule platform and the same annotation process. The authors correctly call the evidence preliminary and say multicentre prospective validation is required before deployment.[1]
How 814 patients became two different datasets
The source database covered 814 patients who underwent small-bowel capsule endoscopy from January 2014 to October 2025. After cleaning, 201,860 images remained. Three experienced readers annotated lesions, with a mean of ten years' experience and more than 500 capsule studies reviewed per year on average. The lesion-recognition dataset included 725 patients: 507 for training and 218 for testing, with ulcers, oedema and stenosis represented in each split.
The much harder scoring task used a smaller subset of 89 patients who had expert-consensus Lewis scores. Forty-nine trained the score model and 40 formed the internal test. The readers' agreement before consensus was high, with an intraclass correlation coefficient of 0.91. Patient-level splitting reduces leakage from repeated frames belonging to the same examination, but the narrow scoring denominator still leaves considerable risk that performance depends on this centre's case mix, readers and acquisition protocol.[1]
The model watches a structured sample, not the full video
A capsule examination can contain tens of thousands of frames. The researchers first resampled each small-bowel sequence at one frame per second, yielding roughly 12,600 frames for a typical 3.5-hour transit. They divided the sequence into proximal, middle and distal thirds and selected 300 contiguous frames around the midpoint of each third. The final 900-frame input represented about 7.1% of the one-frame-per-second sequence. During training the window centres moved slightly; inference used fixed central windows.
That strategy makes sequence modelling computationally manageable and preserves broad anatomical order. It can also miss lesions outside the sampled windows. Crohn disease can be patchy, and a model that estimates a whole-examination score from selected intervals needs evidence that clinically important ulcers or strictures are not systematically omitted. The paper reports lesion-level accuracy and score agreement but no patient-level miss rate comparing sampled and exhaustive review of the same videos.[1]
What the performance numbers show
Across ulcer, oedema and stenosis recognition, the model's mean average precision was 0.94 with reported inference time of 15 milliseconds per frame. Category recall was 76.7% for ulcers, 82.1% for oedema and 95.7% for stenosis; precision was 93.8%, 99.2% and 98.8%. High precision means most flagged lesions matched the reference, while lower ulcer recall means almost one in four reference ulcers was not detected in that test.
For the 40-case Lewis-score test, mean absolute error was 2.9, root mean squared error 3.833 and R-squared 0.82, with patient-level bootstrap intervals. Removing the lesion detector, temporal module or stenosis branch worsened performance, supporting the multitask design. These metrics measure agreement with expert reference scores, not improved diagnosis or outcomes. The paper does not report a head-to-head reader-time trial, treatment decisions, flare prediction or whether clinicians make fewer errors with the tool.[1]
Why the later 20 cases do not prove generalisation
The model was applied without retraining to 20 consecutive cases collected in 2026. Mean absolute error was 3.3 and root mean squared error 5.182. R-squared rounded to 1.00, but the higher squared error shows that some absolute deviations were larger than in internal testing. With only 20 cases and a restricted score distribution, a variance-based statistic can look excellent while remaining fragile. The authors explicitly warn against calling this perfect prediction.
Five control examinations with no endoscopic evidence of Crohn-related inflammation received model scores of 1, 3, 3, 1 and 0. Four therefore had a non-zero false-positive score, although all were far below the clinical threshold of 135 for mild disease. This is reassuring only in a very small sample. A useful external study must include normal variants, bowel preparation artefacts, other inflammatory conditions, different ages and multiple capsule manufacturers.[1]
What clinicians and patients should expect next
The near-term role is a review aid: highlighting candidate lesions, offering a draft score and making its sampled frames visible so an endoscopist can accept, correct or reject the output. That could reduce repetitive viewing and improve consistency, but the workflow must preserve access to the full recording and make uncertainty explicit. A low automated score should not close an investigation when symptoms, biomarkers or another test suggest active disease.
Confidence would rise with preregistered external validation across hospitals, devices and independent readers, followed by a prospective assistance trial. That trial should measure missed clinically important lesions, review time, inter-reader agreement, changes in treatment, downstream investigations and patient outcomes. One author is employed by the capsule manufacturer, while clinical annotation was performed by the hospital team; the relationship is disclosed and makes independent replication especially important. The study advances the engineering case, not the case for unsupervised clinical use.[1]
What this means for people
- People with Crohn disease may eventually receive faster, more consistent capsule-video review, but the current system has not been shown to improve diagnosis, treatment or health outcomes.
- Endoscopists need the full video, visible model evidence and an easy correction path because the sequence model samples only part of a typical recording and ulcer recall was 76.7%.
- A hospital considering adoption should validate the model on its own capsule devices, preparation quality, patient mix and readers before using scores in care.
Global context
Capsule endoscopy is used internationally, but equipment, bowel preparation, disease prevalence, reading practice and access to specialist review differ. This dataset came from one Chinese children's hospital and one manufacturer's system. The evidence therefore cannot establish performance in adult services, other countries, lower-resource clinics or different capsule platforms.
What the evidence does not yet show
- The Lewis-score model trained on 49 cases and was internally tested on 40; the later cohort contained only 20 cases from the same centre.
- The model analysed 900 selected frames, about 7.1% of a typical one-frame-per-second sequence, rather than exhaustively processing the full recording.
- There was no multicentre, cross-device or independently annotated external validation.
- Ulcer recall was 76.7%, and only five control examinations were tested for specificity.
- The study did not measure clinician time, assisted-reader accuracy, treatment decisions, patient outcomes or live deployment failures.
What to watch next
- External validation across adult and paediatric centres and multiple capsule systems
- Full-video or adaptive-sampling comparisons that quantify missed lesions
- Prospective trials of clinician-plus-AI review, including time, errors and treatment changes
- Independent replication and clear management of the disclosed manufacturer relationship
Living evidence record
Impact record IAI-0HH08ZR
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
14 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what JMIR Medical Informatics published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 14 October 2026
Research question, topics and regions
Whether a multitask deep-learning system can detect small-bowel Crohn lesions and estimate an examination-level Lewis score from sampled capsule-endoscopy sequences
Topics: Crohn disease · Capsule endoscopy · Medical imaging · Deep learning · Clinical scoring · External validation
Regions: China · Asia & Middle East · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Health & Life Sciences
Can brain MRI reveal more than BMI?
A peer-reviewed study trained a deep-learning model on 45,702 MRI scans from six cohorts. Brain-derived features tracked BMI and separated several disease groups better than BMI alone—but the disease analysis stayed inside UK Biobank and cannot establish cause, diagnosis or clinical benefit.
9 min · 2 sources
Health & Life Sciences
An Alzheimer’s MRI model scored 94.5%. On new data, it fell to 61.6%
For a patient, the important test is whether a model works beyond familiar images. This classifier’s accuracy dropped sharply on an external dataset, and it remains a research system rather than a clinical diagnostic tool.
7 min · 1 source
Health & Life Sciences
Can AI spot liposarcoma on ultrasound?
A peer-reviewed Chinese study reports strong results in a 95-patient internal test, including 0.97 accuracy. But all 317 patients came from one hospital and there was no external validation, so this is a proof of concept—not a clinically cleared diagnostic system.
9 min · 1 source
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.