Scholarlink

<p>Artificial intelligence (AI)-based digital morphology analyzers are vital for clinical hematology but face critical challenges in identifying immature granulocytes and atypical cells. This study validated the MC-100i AI system against 15 morphologists, with an additional 3 senior experts establishing a gold standard, to assess performance gaps and clinical utility. A total of 104 blood smears containing 19,174 cells (9 abnormal types, including 3154 leukocytes and nucleated erythrocytes) were analyzed. The AI achieved an overall accuracy of 95.97% (ranked second) and 91.38% accuracy for the abnormal subsets (ranked fifth). Critical biases were identified: AI classified ambiguous cells (e.g., promyelocytes) as earlier developmental stages, while humans favored later stages. AI relied on isolated features for atypical cells, unlike experts who integrated smear context. These findings confirm the AI’s robust clinical potential for routine leukocyte classification, and targeted optimizations will further enhance AI-driven peripheral blood cell identification for effective integration into clinical diagnostic workflows.</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

Evaluating AI in leukocyte classification: performance of the AI system against 15 morphology experts

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月11日

查看全文
  • Wei Xu,
  • Haoqin Jiang,
  • Jingxian Zhang,
  • Kun Chen,
  • Yang Fei,
  • Ping Guo,
  • Yongjian He,
  • Yi Ye,
  • Linlin Qu,
  • Mingkang Yang,
  • Lizhi Yan,
  • Di Wang,
  • Huan Qi,
  • Shihong Zhang,
  • Chi Zhang,
  • Jianbiao Wang,
  • Ming Guan

摘要

Artificial intelligence (AI)-based digital morphology analyzers are vital for clinical hematology but face critical challenges in identifying immature granulocytes and atypical cells. This study validated the MC-100i AI system against 15 morphologists, with an additional 3 senior experts establishing a gold standard, to assess performance gaps and clinical utility. A total of 104 blood smears containing 19,174 cells (9 abnormal types, including 3154 leukocytes and nucleated erythrocytes) were analyzed. The AI achieved an overall accuracy of 95.97% (ranked second) and 91.38% accuracy for the abnormal subsets (ranked fifth). Critical biases were identified: AI classified ambiguous cells (e.g., promyelocytes) as earlier developmental stages, while humans favored later stages. AI relied on isolated features for atypical cells, unlike experts who integrated smear context. These findings confirm the AI’s robust clinical potential for routine leukocyte classification, and targeted optimizations will further enhance AI-driven peripheral blood cell identification for effective integration into clinical diagnostic workflows.