We also have X and podcasts
AI Rivals Expert Morphologists in Leukocyte Classification
An AI system matched expert-level performance in leukocyte classification, achieving 95.97% accuracy and ranking second among 16 participants.
Background
Accurate leukocyte classification from blood smears is critical for diagnosing hematologic disorders, but demands expertise and time. Researchers compared the Mindray MC-100i AI system against 15 morphology experts, analyzing 19,174 cells from 104 blood smears across 14 leukocyte types and nucleated erythrocytes.
Key Findings
- AI achieved 95.97% accuracy overall, excelling on normal leukocytes (98.57%) but showing lower accuracy on abnormal cells (91.38%)
- Eosinophil identification reached 99.65% accuracy, surpassing all human experts
- Abnormal promyelocyte detection achieved 94.65%—clinically valuable for early Acute Promyelocytic Leukemia diagnosis
- AI systematically classified ambiguous cells as earlier developmental stages, while humans favored later stages
- Immature granulocytes and atypical cells presented the greatest classification challenges
Why It Matters
The AI’s conservative strategy—classifying ambiguous cells as earlier developmental stages—offers clinical value for time-sensitive conditions like APL. It shows promise as a complementary screening tool, reducing expert workload while maintaining high accuracy.
Limitations
Performance gaps were most pronounced for morphologically complex cells. The AI analyzes isolated features, while experts integrate contextual smear information. Further optimization for immature granulocytes and atypical cells is needed for clinical integration.
Original paper: Evaluating AI in leukocyte classification: performance of the AI system against 15 morphology experts. — NPJ digital medicine. 10.1038/s41746-026-02601-w
