Accuracy of Automatic Quantitative Analysis of Pathological Tissue Images Using ChatGPT Data Analyst: Comparison With...

Colorectal cancer has a high incidence and mortality rate worldwide. In advanced cases, liver metastasis frequently occurs and determines the prognosis. Image analysis of pathological tissues from organs with metastasis is essential for tumor research and diagnosis, but relies on manual or semi-automated software processing, challenging user expertise, and efficiency. We evaluated the image analysis capabilities of ChatGPT (Data Analyst) to automatically quantify tumor regions in H&E-stained sections from a mouse metastasis model established using MC38 cells. Using ChatGPT Data Analyst, we analyzed the tumor area, total liver area, tumor proportion, and tumor count and compared the results with those obtained using conventional ImageJ analysis. Area measurements showed a high correlation (r > 0.99) and agreement (ICC > 0.9) between the ChatGPT Data Analyst and ImageJ, with high reproducibility upon reanalysis. Given that the clarity of the boundary between the tumor and liver parenchyma is influenced by staining conditions, we investigated optimal protocols and identified specific conditions (hematoxylin for 2 min, eosin for 30 s) that are crucial for enhancing extraction accuracy. Furthermore, comprehensive benchmarking revealed that advanced general-purpose LLMs (GPT-5.1 and Gemini-2.5 Pro) achieved accuracy comparable to that of a specialized AI framework (SAM + PLIP) and commercial software (NIS-Elements AR). ChatGPT Data Analyst offers a novel method for pathological image analysis based on prompt instructions, with an accuracy comparable to that of conventional methods. Future improvements in tumor count detection through ChatGPT Data Analyst updates and standardized staining conditions will enhance its application in pathological research and diagnosis of tumors.