Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 25 of the 26 references it cites were found in a registry or an index.

Exploring the risks of over-reliance on AI in diagnostic pathology. What lessons can be learned to support the training of young pathologists?

Open access
Bellahsen-Harrar, Yaëlle, Lubrano, Mélanie, Lépine, Charles, Beaufrère, Aurélie, Bocciarelli, Claire, Brunet, Anaïs, Decroix, Elise, El-Sissy, Franck Neil, Fabiani, Bettina, Morini, Aurélien, Tilmant, Cyprien, Walter, Thomas, Badoual, Cécile · 2025 · PLOS One

Abstract

The integration of Artificial Intelligence (AI) algorithms into pathology practice presents both opportunities and challenges. Although it can improve accuracy and inter-rater reliability, it is not infallible and can produce erroneous diagnoses, hence the need for pathologists to always check predictions. This critical judgment is particularly important when algorithm errors could lead to high-impact negative clinical outcomes, such as missing an invasive carcinoma. However, the influence of AI tools on pathologists’ decision-making is not well explored. This study aims to evaluate the impact of a previously developed AI tool on the diagnostic accuracy and inter-rater reliability among pathologists, while assessing whether pathologists maintain independent judgment of AI predictions. Eight pathologists from different hospitals and with varying levels of experience, participated in the study. Each of them reviewed 115 slides of laryngeal biopsies, including benign epithelium, low-grade and high-grade dysplasia, and invasive squamous carcinomas. The study compared diagnostic outcomes with and without AI assistance. The reference labels were established by an expert’s double-blind review. Results show that assisted pathologists had a higher accuracy for high-grade dysplasia, invasive carcinoma and improved inter-rater reliability. However, cases of over-reliance on AI have been observed, resulting in the omission of correctly diagnosed invasive carcinomas during the unassisted examination. The false predictions on these carcinoma slides were labeled with a low confidence score, which was not considered by the less experienced pathologists, showing the risk that they would follow the AI prediction without enough critical judgment or expertise. Our study emphasizes the potential over-reliance of pathologists on AI models and the potential harmful consequences, even with the advancement of powerful algorithms. The integration of confidence scores and the education of pathologists to use this tool could help to optimize the safe integration of AI into pathology practice.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1371/journal.pone.0323270

Cites

25 of 26 references verified1 citation errors · 0 not checkable

Statistics

Figures

Fig 1 Assisted review on the digital platform. (A) Review protocol. The pathologists began either with assisted or non-assisted review and switched after a washout period of at least two weeks. (B) A whole slide image viewer. displaying for each case the virtual slide, the model’s prediction, the confidence score expressed as a percentage, a categorization of the confidence (high or low) and a heatmap highlighting regions of the slide that contributed to the prediction.
Fig 1 Assisted review on the digital platform. (A) Review protocol. The pathologists began either with assisted or non-assisted review and switched after a washout period of at least two weeks. (B) A whole slide image viewer. displaying for each case the virtual slide, the model’s prediction, the confidence score expressed as a percentage, a categorization of the confidence (high or low) and a heatmap highlighting regions of the slide that contributed to the prediction.CC BY · PubMed Central
Fig 2 Digital web-based review platform. (A) Slide information. Framed in blue: model’s prediction, confidence score (%) and confidence level (low or high). (B) Virtual slide. (C) Overlaid heatmap showing in red the tiles that drove the model’s attention and in blue the tiles with a low attention score.
Fig 2 Digital web-based review platform. (A) Slide information. Framed in blue: model’s prediction, confidence score (%) and confidence level (low or high). (B) Virtual slide. (C) Overlaid heatmap showing in red the tiles that drove the model’s attention and in blue the tiles with a low attention score.CC BY · PubMed Central
Fig 3 Agreement between pathologists with and without AI-assistance. (A) Overall linear kappa. These plots show the linear kappa values for each pathologist with and without the AI-Assistance, compared to the standalone AI-model. The assisted review improved the inter-rater agreement and drastically reduced the kappa’s range between pathologists. (B) Pairwise agreement. This figure illustrates the increase of the inter-rater agreement without considering the reference standard labels (linear kappa 0.734 assisted versus 0.619 non-assisted).
Fig 3 Agreement between pathologists with and without AI-assistance. (A) Overall linear kappa. These plots show the linear kappa values for each pathologist with and without the AI-Assistance, compared to the standalone AI-model. The assisted review improved the inter-rater agreement and drastically reduced the kappa’s range between pathologists. (B) Pairwise agreement. This figure illustrates the increase of the inter-rater agreement without considering the reference standard labels (linear kappa 0.734 assisted versus 0.619 non-assisted).CC BY · PubMed Central
Fig 4 Examples of disagreements between AI model and pathologists of low clinical impact. Magnification from left to right: x2, x10, x20. A. slide_1, a normal epithelium wrongly diagnosed as low-grade dysplasia, which had a tangent inclusion potentially difficult to interpret. B. slide_2, a normal epithelium wrongly diagnosed as low-grade dysplasia, with a tangent cut of the basal layers which can be misleading.
Fig 4 Examples of disagreements between AI model and pathologists of low clinical impact. Magnification from left to right: x2, x10, x20. A. slide_1, a normal epithelium wrongly diagnosed as low-grade dysplasia, which had a tangent inclusion potentially difficult to interpret. B. slide_2, a normal epithelium wrongly diagnosed as low-grade dysplasia, with a tangent cut of the basal layers which can be misleading.CC BY · PubMed Central
Fig 5 Wrongly predicted carcinomas missed by the pathologists during the assisted reviews. Magnification from left to right: x2, x10, x20. A. Slide_3: sometimes correctly diagnosed during the non assisted review, this slide shows tangent cuts making the prediction more difficult for the AI model. B. Slide_4: correctly diagnosed during the non assisted review, this slide shows invasive carcinoma under a normal epithelium, which is an unusual presentation. C. Slide_5: correctly diagnosed during the non assisted review, this slide shows crushing artifacts, which might impair the prediction. D. Slide_6: missed during the non assisted review, this slide shows tangent cuts difficult to interpret. E. Slide_7: missed during the non assisted review, this slide also shows tangent cuts. F. Slide_8: sometimes correctly diagnosed during the non assisted review, this slide shows tangent cuts.
Fig 5 Wrongly predicted carcinomas missed by the pathologists during the assisted reviews. Magnification from left to right: x2, x10, x20. A. Slide_3: sometimes correctly diagnosed during the non assisted review, this slide shows tangent cuts making the prediction more difficult for the AI model. B. Slide_4: correctly diagnosed during the non assisted review, this slide shows invasive carcinoma under a normal epithelium, which is an unusual presentation. C. Slide_5: correctly diagnosed during the non assisted review, this slide shows crushing artifacts, which might impair the prediction. D. Slide_6: missed during the non assisted review, this slide shows tangent cuts difficult to interpret. E. Slide_7: missed during the non assisted review, this slide also shows tangent cuts. F. Slide_8: sometimes correctly diagnosed during the non assisted review, this slide shows tangent cuts.CC BY · PubMed Central
Fig 6 Pathologists’ performances depending on the confidence level. The model’s confidence score guided the pathologists and improved the inter-rater agreement, with higher linear kappa on confident predictions (high confidence predictions: linear kappa 0.8 assisted versus 0.73 non-assisted, p < 0.001).
Fig 6 Pathologists’ performances depending on the confidence level. The model’s confidence score guided the pathologists and improved the inter-rater agreement, with higher linear kappa on confident predictions (high confidence predictions: linear kappa 0.8 assisted versus 0.73 non-assisted, p < 0.001).CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 26 references

Research works printed but not linked: 1

  1. 2017, WHO classification of head and neck tumours

Cited by

See all 14

Metadata

Authors
Bellahsen-Harrar, Y., Lubrano, M., Lépine, C., Beaufrère, A., Bocciarelli, C., Brunet, A., Decroix, E., El-Sissy, F. N., Fabiani, B., Morini, A., Tilmant, C., Walter, T., & Badoual, C.
Affiliations
  • Université Paris Sciences et Lettres (FR)
  • Inserm (FR)
  • Centre Hospitalier Régional Universitaire de Brest (FR)
  • Université Paris-Est Créteil (FR)
  • Institut universitaire du cancer de Toulouse Oncopole (FR)
  • Hôpital Lariboisière (FR)
  • Groupe Hospitalier de l'Institut Catholique de Lille (FR)
  • Université Paris Cité (FR)
Corresponding author
Badoual, Cécile
Venue
PLOS One
Published
2025-08-28
Volume
20
Issue
8
Pages
e0323270
ISSN
1932-6203 (ISSN-L)
OpenAlex
W4413814674
Cited by
11
References
26
Citations by year
2026: 10 · 2025: 1
FWCI
2.63
Type
journal article
Language
en
Open access
Locations
  • PLoS ONE (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • HAL (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
ANRT
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Artificial Intelligence › AI in cancer detection
  • Health Sciences › Medicine › Family Practice › Clinical Reasoning and Diagnostic Skills
Keywords
artificial intelligence in pathology, overreliance on AI, diagnostic accuracy, inter-rater reliability, laryngeal biopsy, squamous cell carcinoma, dysplasia grading, AI-assisted diagnosis, confidence scores, critical judgment, pathologist training, clinical implementation of AI
MeSH terms
Pathologists, Algorithms, Artificial Intelligence, Diagnostic Errors, Humans, Pathology, Clinical, Reproducibility of Results, Observer Variation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Exploring the risks of over-reliance on AI in diagnostic pathology. What lessons can be learned to support the training of young pathologists? https://app.cochat.ai/papers/10.1371/journal.pone.0323270
The paper
Bellahsen-Harrar, Y., Lubrano, M., Lépine, C., Beaufrère, A., Bocciarelli, C., Brunet, A., Decroix, E., El-Sissy, F. N., Fabiani, B., Morini, A., Tilmant, C., Walter, T., & Badoual, C. (2025). Exploring the risks of over-reliance on AI in diagnostic pathology. What lessons can be learned to support the training of young pathologists? PLOS One. https://doi.org/10.1371/journal.pone.0323270

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.