Scholarlink

<p>Microscopic morphology in clinical laboratories is expertise-intensive, variable across observers, and difficult to scale. We present Lingjian, a vision-language model tailored for laboratory morphology. Built on Qwen3-VL-8B and trained via multistage domain adaptation on more than 400,000 laboratory images with paired text and grounding supervision, Lingjian unifies image comprehension, cell identification, morphology description, and interpretation, and target localization. Across public benchmarks and cross-domain test sets, Lingjian shows robust multitask performance. On China’s National Center for Clinical Laboratories External Quality Assessment (EQA, 2021–2025), it reached 93.0% overall accuracy, outperforming a human expert group (78.1%) and strong general models (e.g., Gemini-3 Pro, 75.3%). In a 120-case multireader multicase reader study, Lingjian's assistance improved junior-reader sensitivity for abnormal screening from 69.7% to 91.7% while maintaining high specificity (85.0% to 89.5%). Model weights and evaluation resources are released to support reproducibility.</p>

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

An expert-level vision-language model for multitask diagnostic morphology in clinical laboratories

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Han Liu,
  • Yuhan Zhu,
  • Shunbo Li,
  • Qinli Pu,
  • Liping Yang,
  • Weixian Chen,
  • Juan Hu

摘要

Microscopic morphology in clinical laboratories is expertise-intensive, variable across observers, and difficult to scale. We present Lingjian, a vision-language model tailored for laboratory morphology. Built on Qwen3-VL-8B and trained via multistage domain adaptation on more than 400,000 laboratory images with paired text and grounding supervision, Lingjian unifies image comprehension, cell identification, morphology description, and interpretation, and target localization. Across public benchmarks and cross-domain test sets, Lingjian shows robust multitask performance. On China’s National Center for Clinical Laboratories External Quality Assessment (EQA, 2021–2025), it reached 93.0% overall accuracy, outperforming a human expert group (78.1%) and strong general models (e.g., Gemini-3 Pro, 75.3%). In a 120-case multireader multicase reader study, Lingjian's assistance improved junior-reader sensitivity for abnormal screening from 69.7% to 91.7% while maintaining high specificity (85.0% to 89.5%). Model weights and evaluation resources are released to support reproducibility.