摘要目的:评价人工智能图像分析技术对骨髓细胞形态的识别效能。方法:回顾性研究。2019年12月1日至2020年12月31日在河北医科大学第二医院就诊患者的骨髓标本。选取骨髓标本100份，患者来源包括慢性髓系细胞白血病23例、骨髓增生异常综合征4例、慢性淋巴细胞白血病8例、多发性骨髓瘤5例、急性白血病7例、慢性贫血患者32例、感染患者6例及正常骨髓象15名。其中男45例，女55例，年龄52（37，66）岁，骨髓涂片经瑞-姬染色后，应用AI分析系统和人工审核对13种骨髓有核细胞进行分类，以人工审核结果为金标准，比较两种方法的结果差异，应用统计学软件绘制混淆矩阵，人工审核结果与AI分析系统预分类结果符合度采用Kappa一致性检验方法进行统计；应用Pearson检验分析AI系统预分类和手工镜检分类结果的一致性。选取临床已确诊的MDS和AML骨髓标本各30例，比较AI分析系统人工审核与手工镜检2种方法分类计数原始细胞百分比结果的差异性，以评估AI分析系统的临床应用价值。结果:应用AI分析系统扫描共获得76 630张13种骨髓有核细胞的高清单一彩色图像；13种骨髓有核细胞分类计数的加权平均实验诊断效率参数为：敏感度95.82%，特异度99.19%，准确度98.89%，假阳性率81.00%，假阴性率4.18%；AI分析系统预分类结果和手工镜检分类结果的相关性显示：原始细胞、早幼粒细胞、中性中幼粒细胞、中性晚幼粒细胞、中性杆状核粒细胞、中性分叶核粒细胞、嗜酸性粒细胞、嗜碱性粒细胞、中幼红细胞、晚幼红细胞及淋巴细胞均具有较好的正相关性（ r>0.70， P<0.001）；AI分析系统与手工镜检2种方法对MDS原始细胞的计数结果差异无统计学意义（ P>0.05），对AML原始细胞计数差异有统计学意义（ P<0.01），但由于原始细胞计数均≥20%，符合AML的分型诊断标准。 结论:AI分析系统对13种骨髓有核细胞分类计数有较好的敏感度、特异度和准确度。

abstractsObjective:To evaluate the screening efficacy of AI for bone marrow cell morphology.Method:Bone marrow specimens of patients attending the Second Hospital of Hebei Medical University from December 1,2019 to December 21,2020;(1) Selected from one hundred bone marrow specimens, The cases included chronic myeloid cell leukemia ( n=23), myelodysplastic syndrome ( n=4), chronic lymphocytic leukemia ( n=4), multiple myeloma ( n=5), 7 acute leukemia ( n=7), chronic anemia ( n=32), infection ( n=6) and healthy control ( n=15). Including 45 males and 55 females, with age 52(37,66)years old.The bone marrow smear prepared with Wright-Giemsa, The AI analysis system and manual audit were applied to classify 13 types of bone marrow nucleated cell, taking the results of manual audit as the gold standard, comparing the difference between the results of the two methods, using statistical software to draw the confusion matrix, The compliance between the manual audit results and the pre-classification results of the AI analysis system was calculated by the Kappa consistency test method; The consistency analysis between the pre-classification results of AI and those of the manual microscopic examination was performed by the Pearson test; (2)Statistics analyzed the blast cell differential count differences of AI and manual microscopy, to evaluate the clinical application value of AI analysis system, which soured from thirty bone marrow samples of patients diagnosed with MDS and AML. Results:76 630 images of 13 nucleated cells were obtained by AI analysis system; the weighted average experimental diagnostic efficiency parameters of 13 types of bone marrow nucleated cells, are as follows: sensitivity(%)=95.82, specificity(%)=99.19, accuracy(%)=98.89, false positive rate(%)=0.81, false negative rate (%)=4.18; the correlation results, between the pre-classification results of AI and manual microscopic classification results,showed that blast cell, promyelocytes, neutrophilic myelocyte, neutrophilic metamyelocyte, band neutrophil, segmented neutrophi,eosinophil, basophil, polychromatic erythroblast, orthochromatic erythroblast, and lymphocytes have good positive correlation ( r>0.70,all P<0.001), while basophilic erythroblast and monocytes have no obvious correlation ( r=0.32,0.30, all P> 0.001); the count results of the blast cells in bone marrow smears of MDS and AML, got by AI and manual microscopy respectively, showed that the average percentage of blast cells was 8.19% by AI and 8.68% by manual microscopy in MDS, there was no significant difference between the two methods ( P>0.05); the average percentage of blast cells was 48.52% by AI analysis system and 53.77% by manual microscopy in AML, and although there was a significant difference in blast cell count ( P<0.01), coincidence the classification diagnostic criteria for AML (blast cells ≥ 20%). Conclusion:The AI analysis system performed good sensitivity, specificity and accuracy for 13 types of bone marrow nucleated cells, which showed potential application value for the rapid classification and diagnosis of MDS and AML.