Illusion of competence: vision-language models provide confident but inaccurate explanations in cytological diagnostics.

Ivan Kukuljan¹, Muhammed Furkan Dasdelen¹, Julia Schäfer^1,2, Michele Buck^3,4, Katharina S Götze^3,4, Carsten Marr^5,6,7,8,9

¹Computational Health Center & Helmholtz AI, Helmholtz Munich, Neuherberg, Germany.

²School of Medicine and Health, Technical University of Munich, Munich, Germany.

³Department of Medicine III, Hematology/Oncology, Technical University of Munich School of Medicine and Health, Munich, Germany.

⁴Bavarian Center for Cancer Research (BZKF), Munich, Germany.

⁵Computational Health Center & Helmholtz AI, Helmholtz Munich, Neuherberg, Germany. carsten.marr@helmholtz-munich.de.

⁶Department of Medicine III, LMU Medizin, LMU Munich, Munich, Germany. carsten.marr@helmholtz-munich.de.

⁷Department of Physics, Ludwig-Maximilians-Universität München, Munich, Germany. carsten.marr@helmholtz-munich.de.

⁸German Cancer Consortium (DKTK), partner site Munich, Germany. carsten.marr@helmholtz-munich.de.

⁹Munich Center for Machine Learning (MCML), Munich, Germany. carsten.marr@helmholtz-munich.de.

PubMed PMC DOI PDF

Abstract

Large vision-language models (LVLMs) have shown impressive image-understanding capabilities across domains. However, their suitability for cytomorphological diagnostics remains unclear. Here, we systematically evaluated four state-of-the-art generalist LVLMs, GPT-4o, Gemini-2.0, Llama-3.2, and DeepSeek-VL2, and three biomedical LVLMs, LLaVA-Med, CONCH, and BiomedCLIP, across key cytomorphology ben…

Create a free account to keep reading

Free members get 10 full research views every month across publications, clinical trials, FDA clearances, adverse events, and NIH grants. No credit card required.

Create Free Account Sign In

Want unlimited research access? See Pro plans

Data Accuracy Notice: Research intelligence on Health AI Central is aggregated from public sources (PubMed, ClinicalTrials.gov, FDA, NIH, CMS, and others) and refreshed nightly. Classifications and derived metrics are produced by automated methods described in our Methodology. We recommend verifying critical data points against the primary sources before making decisions.