A multi-center study on evaluation of leukocyte differential performance by an artificial intelligence-based Digital Cell Morphology Analyzer / 中华检验医学杂志

ABSTRACT

Objective:To evaluate the performance of an artificial intelligent (AI)-based automated digital cell morphology analyzer (hereinafter referred as AI morphology analyzer) in detecting peripheral white blood cells (WBCs).Methods:A multi-center study. 1. A total of 3010 venous blood samples were collected from 11 tertiary hospitals nationwide, and 14 types of WBCs were analyzed with the AI morphology analyzers. The pre-classification results were compared with the post-classification results reviewed by senior morphological experts in evaluate the accuracy, sensitivity, specificity, and agreement of the AI morphology analyzers on the WBC pre-classification. 2. 400 blood samples (no less than 50% of the samples with abnormal WBCs after pre-classification and manual review) were selected from 3 010 samples, and the morphologists conducted manual microscopic examinations to differentiate different types of WBCs. The correlation between the post-classification and the manual microscopic examination results was analyzed. 3. Blood samples of patients diagnosed with lymphoma, acute lymphoblastic leukemia, acute myeloid leukemia, myelodysplastic syndrome, or myeloproliferative neoplasms were selected from the 3 010 blood samples. The performance of the AI morphology analyzers in these five hematological malignancies was evaluated by comparing the pre-classification and post-classification results. Cohen′s kappa test was used to analyze the consistency of WBC pre-classification and expert audit results, and Passing-Bablock regression analysis was used for comparison test, and accuracy, sensitivity, specificity, and agreement were calculated according to the formula.Results:1. AI morphology analyzers can pre-classify 14 types of WBCs and nucleated red blood cells. Compared with the post-classification results reviewed by senior morphological experts, the pre-classification accuracy of total WBCs reached 97.97%, of which the pre-classification accuracies of normal WBCs and abnormal WBCs were more than 96% and 87%, respectively. 2. The post-classification results reviewed by senior morphological experts correlated well with the manual differential results for all types of WBCs and nucleated red blood cells (neutrophils, lymphocytes, monocytes, eosinophils, basophils, immature granulocytes, blast cells, nucleated erythrocytes and malignant cells r>0.90 respectively, reactive lymphocytes r=0.85). With reference, the positive smear of abnormal cell types defined by The International Consensus Group for Hematology, the AI morphology analyzer has the similar screening ability for abnormal WBC samples as the manual microscopic examination. 3. For the blood samples with malignant hematologic diseases, the AI morphology analyzers showed accuracies higher than 84% on blast cells pre-classification, and the sensitivities were higher than 94%. In acute myeloid leukemia, the sensitivity of abnormal promyelocytes pre-classification exceeded 95%. Conclusion:The AI morphology analyzer showed high pre-classification accuracies and sensitivities on all types of leukocytes in peripheral blood when comparing with the post-classification results reviewed by experts. The post-classification results also showed a good correlation with the manual differential results. The AI morphology analyzer provides an efficient adjunctive white blood cell detection method for screening malignant hematological diseases.