Scholarlink

This study delivers the first multi-vendor comparison of closed source Vision-Language Models (VLMs) for automated Ki-67 index estimation in breast cancer histopathology. This study applies a single guideline-based prompt to seven proprietary VLMs to identify Ki-67 positive and negative nuclei, compute the proliferation index, and show the calculation steps. On the expert-annotated BCData test set (402 images), performance varies substantially: GPT-4.5 achieves the highest concordance with pathologists ( \(R^{2}=0.86\) , RMSE = 7.97), while Gemini 1.5 Pro and Grok-2 Vision score lower ( \(R^{2}=0.62\) and 0.28, respectively). Inference time per image ranged from 1.5 to 6 s, reflecting different speed-cost trade-offs. This study shows that closed VLMs can estimate Ki-67 without retraining at a level that may be clinically useful; however, accuracy and cost vary by provider.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative Study of Closed Vision-Language Models for Ki-67 Index Prediction in Breast Cancer Histopathology Images

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Harold Brayan Arteaga-Arteaga,
  • Juan Andrés Giraldo-Arias,
  • Esteban Mercado-Ruiz,
  • Miguel Angel Londoño-Salgado,
  • Cristina Juárez Landín,
  • Mario Alejandro Bravo-Ortiz,
  • Pablo Guillen-Rondon,
  • Reinel Tabares-Soto

摘要

This study delivers the first multi-vendor comparison of closed source Vision-Language Models (VLMs) for automated Ki-67 index estimation in breast cancer histopathology. This study applies a single guideline-based prompt to seven proprietary VLMs to identify Ki-67 positive and negative nuclei, compute the proliferation index, and show the calculation steps. On the expert-annotated BCData test set (402 images), performance varies substantially: GPT-4.5 achieves the highest concordance with pathologists ( \(R^{2}=0.86\) , RMSE = 7.97), while Gemini 1.5 Pro and Grok-2 Vision score lower ( \(R^{2}=0.62\) and 0.28, respectively). Inference time per image ranged from 1.5 to 6 s, reflecting different speed-cost trade-offs. This study shows that closed VLMs can estimate Ki-67 without retraining at a level that may be clinically useful; however, accuracy and cost vary by provider.