Publikationsdatenbank

Does artificial intelligence influence pathologists' decisions?
a study on the anchoring bias in computer-assisted mitotic figure assessment by veterinary pathologists (2026)

Art

Poster

Autoren

Conrad, T. (WE 12)

Banerjee, S.

Rosbach, E.

Doherr, M. G. (WE 16)

Bartenschlager, F. (WE 12)

Bolfa, P.

de Brot, S.

Bünger, M. (WE 12)

Donovan, T. A.

Eddicks, L.

Fiedler, S.

Gager, A. M.

Haake, A. F. H. (WE 12)

Hanselmann, L. K. (WE 12)

Kolb, N.

Landmann, K. (WE 12)

Pischon, H.

Radbruch, M.

Weiss, V.

Ammeling, J.

Kaltenecker, C.

Stroblberger, C.

Aubreville, M.

Betram, C. A.

Klopfleisch, R. (WE 12)

Kongress

ESVP/ECVP/ESVCP/ECVCP Congress 2026

Nantes, France, 16. – 19.09.2026

Quelle

Annual ESVP/ECVP/ESVCP/ECVCP Congress 2026 : Proceedingsbook

— S. 288–289

Sprache

Englisch

Verweise

URL (Volltext): https://www.esvp-ecvp-estp-congress.eu/resources/media/ESVP_2026/20260901_ESVP-ECVP-ESVCP-ECVCP_Proceedingsbook_2026.pdf

Kontakt

Institut für Veterinär-Epidemiologie und Biometrie

Königsweg 67
14163 Berlin
+49 30 838 56034
epi@vetmed.fu-berlin.de

Abstract / Zusammenfassung

Background:
Reliable mitotic figure identification is essential for mitotic count assessment in histological tumour sections but shows substantial observer variability. Artificial intelligence (AI) has been shown to improve reproducibility, but AI-generated suggestions may bias pathologists' interpretation. This study evaluated whether veterinary pathologists shift their judgements toward displayed AI-attributed confidence scores when assessing mitotic figures in canine cutaneous mast cell tumours.
Materials & Methods:
Fourteen pathologists scored 500 H&E image patches, each containing a mitotic figure candidate, over three rounds on a 0–100 scale. Scores >50 indicated a mitotic figure and scores <50 indicated no mitotic figure. Two unaided rounds assessed intra-rater variability. In the final round, AI-attributed confidence scores were displayed with the images. These scores were not calculated by a real AI model but were based on each participant's previous ratings and systematically shifted by +/-15, +/-25, or +/-35 points.
Results:
The two unaided rounds showed moderate agreement (intraclass correlation coefficient [ICC] = 0.648, 95% confidence interval [CI] 0.634–0.661), and median absolute intraobserver variability was 10 points. In the AI-supported round, participants shifted their ratings significantly toward the displayed AI-attributed values, with larger discrepancies causing larger shifts in the same direction (beta = 0.118, 95% CI 0.101–0.136, p < 0.001). However, this did not increase binary classification switches beyond unaided variability (22.8% vs. 20.3%).
Conclusion:
AI-attributed confidence scores caused a statistically significant but small anchoring effect on veterinary pathologists' assessments. However, the results indicate that the AIassisted approach did not negatively impact the pathologists' ability to interpret mitotic figure morphologies.