Can endoscopy AI travel beyond still images?
A new classifier reached about 94% accuracy in cross-validation and 96% on an aligned external image subset, then narrowly outscored five gastroenterologists on 200 still images. Both datasets came from Norway, and the work did not test live video, workflow or patient outcomes.
Editorial responsibility: The Impact of AI Editorial Desk · Report a factual concern
At a glance
- 1The model was trained and cross-validated on 8,000 balanced Kvasir-v2 still images across eight classes, reaching 94.24% accuracy and 94.23% F1 score over five stratified folds.
- 2Without additional tuning, it reached 95.98% accuracy on 6,563 images selected from HyperKvasir to match the same eight classes, but both public datasets came from Norwegian health services.
- 3On 200 selected still images the model scored 94%, compared with 85% to 93% for five gastroenterologists; the study did not test live endoscopy video, prospective workflow or patient outcomes.
Research topic
Whether one attention-based image classifier can support several gastrointestinal endoscopy classification tasks and retain performance across a second public dataset, synthetic artifacts and a small reader comparison
One classifier produced seven readouts from eight image classes
The authors built HierTAC-Net, a six-stage convolutional network combining inverted residual blocks, attention at three image resolutions and ConvNeXt-style refinement. It assigns each still image to one of eight Kvasir-v2 classes. The researchers then regroup those probabilities into seven outputs, including full diagnosis, anatomical-landmark checks, pathological classification and polyp screening. That is an engineering convenience: several binary or grouped readouts come from one eight-class softmax rather than seven independently trained models.
Training used Kvasir-v2, a public dataset of 8,000 endoscopy images with exactly 1,000 images in each of eight categories. The images originated at Vestre Viken Health Trust in Norway and were annotated by experienced endoscopists. Five-fold stratified cross-validation produced 94.24% accuracy, 94.23% F1 score and a reported 95% confidence interval of roughly plus or minus 1.14 percentage points. Cross-validation reduces dependence on one train-test split, but every fold still comes from the same curated dataset and class balance.[1]
The Impact Brief · Free
Follow the evidence in health & life sciences.
Get a five-minute weekday briefing on what changed, why it matters and where the evidence comes from. Choose the topics you care about.
The external test was larger but geographically close
For an external check, the team mapped eight HyperKvasir categories onto the training labels and selected 6,563 of that dataset's 10,662 labelled images. Without fine-tuning or changing parameters, the model reached 95.98% accuracy and 96.00% F1 score. Class sizes ranged from 201 to 1,028, making this test less artificially balanced than Kvasir-v2. Performance on a separately assembled collection without retraining is stronger evidence than another internal split.
It is not fully independent international validation. HyperKvasir was collected at Baerum Hospital in Norway between 2008 and 2016, while Kvasir-v2 also came from the Norwegian Vestre Viken service. The study selected only classes that could be aligned with the training taxonomy, so its result does not cover every labelled HyperKvasir category or unfamiliar conditions outside that mapping. Differences in hospitals and collection periods help, but scanners, care pathways, patient populations and annotation conventions may still be related.[1]
Five clinicians formed a useful but small comparison
Five gastroenterologists with one, three, five, six and eight years of experience independently classified the same randomized set of 200 stratified Kvasir-v2 still images without clinical metadata. Their accuracies were 85%, 90%, 92%, 89% and 93%. HierTAC-Net scored 94% on that subset. The model therefore exceeded every reader in this exercise, but the margin over the strongest reader was one percentage point—two images out of 200—not evidence that it can replace a gastroenterologist.
The comparison favours a narrow visual classification task. Clinicians ordinarily integrate video, lesion location, procedure context, prior findings, pathology, patient history and the ability to change viewpoint or clean the lens. The model received only selected still frames and answered within the taxonomy it had been trained to recognize. Five readers from one supervised session cannot represent variation across endoscopists, centres or levels of specialist expertise, and the paper does not report a prospective assisted-versus-unassisted reader trial.[1]
Synthetic image damage tested robustness, not real procedural failure
The researchers added compounded blur, uneven illumination and specular highlights at three severity levels. F1 performance declined as images worsened: under compounded blur it fell from 91.25% at the light level to 83.65% at the extreme level, while extreme versions of the other degradations remained roughly between 87.6% and 89.4%. These experiments are valuable because clean benchmark images hide common visual problems and average accuracy can conceal brittle behaviour.
Synthetic transformations remain approximations. The tests did not simulate every source of failure in endoscopy, including bubbles, mucus, fluid occlusion, rapid camera movement, tissue deformation or a lesion that never appears clearly in a selected frame. A classifier can be fast and accurate once it receives a recognizable image yet fail earlier in the pipeline because the frame is missing or the anatomy is outside its categories. Prospective video evaluation needs to count false alerts, missed lesions, processing delays and failures over entire procedures.[1]
Heatmaps and latency are promising, but not deployment evidence
The paper reports Grad-CAM explanation tests with an insertion area under the curve of 0.78 and deletion AUC of 0.22, intended to show that highlighted regions contribute to predictions. It also reports a 110.14 MB model, 2.72 billion floating-point operations and 0.32 milliseconds of GPU inference per 224-by-224 image on the authors' setup. Those measures help compare architectures and suggest real-time processing may be technically possible.
Neither measure shows clinical usefulness by itself. A heatmap can be faithful to a model's internal decision and still focus on a misleading correlate, while performance on one GPU does not include video decoding, frame selection, network transfer, interface rendering or hospital hardware. The authors built a web application, but the article does not report uptime, cybersecurity assessment, regulatory review, usability testing or integration with an endoscopy system. Calling the architecture suitable for deployment is therefore a hypothesis to test, not a completed deployment result.[1]
Funding and dataset structure matter to interpretation
The work spans researchers affiliated with institutions in Malaysia, Qatar and Bangladesh, while the empirical image collections came from Norway. It was funded by a Qatar Research, Development and Innovation Council undergraduate research grant and supported by a first-place award at the Health Tech Hackathon 2025 associated with World Summit AI Qatar and Qatar University. The authors declare no competing interests. Funding does not invalidate the findings, but it should be visible when assessing an engineering claim with a deployment narrative.
The balanced training set also differs from clinical prevalence. Kvasir-v2 contributes exactly 1,000 images per class, whereas real services do not encounter all findings equally and some important conditions are rare. Accuracy in a balanced eight-class benchmark cannot directly predict positive predictive value, workload from false alerts or performance for disease subtypes absent from the taxonomy. The external subset's unequal classes move closer to reality, but it was still chosen to fit the existing labels.[1]
What would change the assessment
The next evidence should be locked-model testing on complete, consecutive procedures from multiple countries and equipment vendors, with no class filtering and with patient-level separation from all development data. A video study should report per-lesion sensitivity, false alerts per procedure, time to detection, latency and performance under poor preparation. Independent pathology or expert consensus should define the reference standard, and demographic or anatomical subgroup results should be reported where sample sizes permit.
A prospective randomized reader-assistance study would show whether the system helps clinicians rather than merely beating a small reader group on static images. It should measure missed lesions, unnecessary biopsies, procedure time, clinician over-reliance and downstream patient outcomes, while documenting device failures and abstentions. Until such evidence exists, HierTAC-Net is a well-evaluated research classifier across two public Norwegian image collections—not a demonstrated replacement for endoscopists or proof of clinical benefit.[1]
What this means for people
- A reliable assistant could reduce image-review workload and help surface lesions, but this study does not show that patients receive earlier or more accurate diagnoses.
- False reassurance from strong benchmark accuracy could be harmful if missing frames, rare findings or unfamiliar equipment are not tested.
- Clinicians would need clear abstentions and manageable alert rates so that assistance does not add distraction or encourage over-reliance.
Global context
The research team spans Malaysia, Qatar and Bangladesh, but both image collections used for performance claims came from Norway. Endoscopy equipment, disease prevalence, preparation quality and clinical workflow vary across regions. International, patient-level prospective validation is needed before these benchmark results can support adoption elsewhere.
What the evidence does not yet show
- Training and external testing used public still-image collections from Norwegian health services; no international or prospective validation was reported.
- The HyperKvasir test selected 6,563 images whose eight labels could be aligned with the training taxonomy, excluding other labelled categories.
- The reader comparison involved five gastroenterologists and 200 curated still images without clinical metadata; the best clinician was one percentage point behind the model.
- Synthetic blur, lighting and glare tests do not cover bubbles, mucus, fluid occlusion, missing frames or the dynamics of live video.
- Balanced benchmark accuracy does not establish predictive value or false-alert workload under real clinical prevalence.
- The study reported no prospective workflow, patient outcome, regulatory, cybersecurity or usability evaluation.
What to watch next
- Locked-model testing on consecutive full procedures from multiple countries, hospitals and equipment vendors.
- Prospective assisted-versus-unassisted endoscopist trials reporting missed lesions and false alerts per procedure.
- Performance on live video, rare conditions and image failures not represented in the eight-class taxonomy.
- Independent clinical, usability, cybersecurity and regulatory evaluation before deployment claims.
Living evidence record
Impact record IAI-1HSNWBB
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
5 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 Scientific Reports 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 5 October 2026
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.
Continue the story
Related reporting
Health & Life Sciences
Does 99% MRI accuracy mean an Alzheimer’s diagnostic is ready?
No. A peer-reviewed model classified images from one 6,400-image Kaggle collection with reported 99% test accuracy, but it has no external clinical validation and the published test counts do not reconcile cleanly.
8 min · 2 sources
Health & Life Sciences
Can a burn-risk model help before the outcome is known?
A peer-reviewed Iranian study reports very high internal discrimination in 2,266 burn admissions. But the model uses hospital stay, ICU stay and infection information accumulated during care, overpredicted mortality in parts of the test set and has not been validated outside one hospital, so it is not an admission score or a deployable clinical tool.
10 min · 1 source
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
Reader commentary
Add evidence, experience or a question
No account is required. Reader notes are published after a brief civility, relevance and safety check; disagreement is welcome.
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.