人工智能辅助骨髓细胞形态学分析在急性髓系白血病诊疗中的应用

Application of artificial intelligence assists bone marrow cytomorphology analysis in the diagnosis and treatment of acute myeloid leukemia

目的探讨人工智能(AI)细胞形态分析系统在急性髓系白血病(AML)形态学诊断及疗效评估中的应用价值。方法收集2021年6月1日至2022年7月31日在中国医学科学院血液病医院住院和门诊的150例初诊或治疗后急性髓系白血病患者骨髓涂片样本进行回顾性分析。其中初诊组50例,包括男28例,女22例,发病年龄43.5(32.3,58.8)岁;治疗后组100例,包括男36例,女64例,发病年龄34.5(23.0,47.0)岁。以形态学专家分析结果作为金标准,评价AI细胞形态分析系统在AML诊疗中对原始细胞识别的准确性、敏感度及特异度。结果初诊组50份样本AI分析原始细胞比例均≥20%,达到AML诊断标准。AI分析原始细胞的准确性为90.3%、敏感度为85.5%、特异度为98.0%。AI与专家分析的原始细胞比例呈正相关(r=0.882,P<0.001)。治疗后组AI分析原始细胞的敏感度为89.7%、特异度为99.2%。AI与专家分析的原始细胞比例呈正相关(r=0.957,P<0.001)。应用AI分析数据判定,该组有8例样本AI对AML疗效评估结果与专家分析不一致。结论AI细胞形态分析系统在AML形态学诊断及疗效评估中对原始细胞识别准确性及敏感度高、特异度好。

Objective To investigate the value of artificial intelligence(AI)cytomorphologic analysis system in the cytomorphological diagnosis and therapeutic evaluation of acute myeloid leukemia(AML).Methods Bone marrow smear samples were collected from 150 patients with newly diagnosed and treated acute myeloid leukemia who were inpatients and outpatients at the Department of Institute of Hematology and Blood Diseases Hospital,Chinese Academy of Medical Sciences and Peking Union Medical College from June 1,2021 to July 31,2022 for retrospective analysis.Among them,there were 50 patients in the newly diagnosed group,including 28 males and 22 females,with the onset age of 43.5(32.3,58.8)years.There were 100 patients in the post-treatment group,including 36 males and 64 females,with the onset age of 34.5(23.0,47.0)years.The results from cytomorphology expert were used as the gold standard and the Python 3.6.7 was used for analysis to evaluate the accuracy,sensitivity,and specificity of the AI cytomorphologic analysis system for blast cell recognition in AML diagnosis and treatment.Results The proportion of blasts in AI analysis of 50 samples in the newly diagnosed group was≥20%,which met the diagnostic criteria of AML.AI analysis of blasts had an accuracy of 90.3%,sensitivity of 85.5%,and specificity of 98.0%.The correlation coefficient between AI and the proportion of blasts analyzed by experts was positively correlated(r=0.882,P<0.001).Meanwhile,in the post-treatment group,the sensitivity and specificity of AI analysis of blasts were 89.7%and 99.2%,respectively.The correlation coefficient between AI and the proportion of blasts analyzed by experts was positively correlated(r=0.957,P<0.001).According to AI analysis data,there are 8 samples in this group whose AI efficacy evaluation results on AML are inconsistent with expert analysis.Conclusion AI cytomorphologic analysis system has high accuracy,sensitivity and specificity for blast cell recognition in AML morphological diagnosis and therapeutic evaluation.