Wang Meifang, Dong Chunxia, Gao Yan, Li Jianlan, Han Mengru, Wang Lijun
Department of Hematology, The Second Hospital of Shanxi Medical University, Taiyuan, China.
Front Oncol. 2022 Apr 14;12:844978. doi: 10.3389/fonc.2022.844978. eCollection 2022.
Bone marrow biopsy is essential and necessary for the diagnosis of patients with aplastic anemia (AA), myelodysplastic syndromes (MDS), and acute myeloid leukemia (AML). However, the convolutional neural networks (CNN) model that automatically distinguished AA, MDS, and AML based on bone marrow smears has not been reported.
Image-net pretrained model of CNN was used to construct the recognition model. Data extracted from the American Society of Hematology (ASH) Image Bank were utilized to develop the model and data extracted from the clinic were used for external validation. The model had two output layers: whether the patient was MDS (two-classification) and which of AA, MDS, and AML the patient was (three-classification). Different outcome weights (two-classification/three-classification = 5:5, 2:8, 1:9) and epochs (30, 50, 200) were used to select the optimal model. The model performance was evaluated by the Accuracy-Loss curves and calculating the area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).
A total of 115 bone marrow smears from the ASH Image Bank and 432 bone marrow smears from the clinic were included in this study. The results of Accuracy-Loss curves showed that the best model training effect was observed in the model with the outcome weight and epoch of 1:9 and 200. Similarly, this model also performed well performances in the two-classification of MDS and the three-classification of AA, MDS, AML. The AUC, accuracy and sensitivity of the MDS two-classification model in the testing set were 0.985 [95% confidence interval (CI), 0.979-0.991], 0.914 (95%CI, 0.895-0.934), and 0.992 (95%CI, 0.980-1.000), respectively. The AUC, accuracy and sensitivity of the AA, MDS, AML three-classification model in the testing set were 0.968 (95%CI, 0.960-0.976), 0.929 (95%CI, 0.916-0.941), and 0.857 (95%CI, 0.828-0.886), respectively.
The image-net pretrained model was able to obtain high accuracy AA, MDS, AML distinction, and may provide clinicians with a convenient tool to distinguish AA, MDS, and AML.
骨髓活检对于再生障碍性贫血（AA）、骨髓增生异常综合征（MDS）和急性髓系白血病（AML）患者的诊断至关重要且必不可少。然而，尚未有基于骨髓涂片自动区分AA、MDS和AML的卷积神经网络（CNN）模型的报道。
使用CNN的Image-net预训练模型构建识别模型。从美国血液学会（ASH）图像库提取的数据用于开发该模型，从临床提取的数据用于外部验证。该模型有两个输出层：患者是否为MDS（二分类）以及患者是AA、MDS和AML中的哪一种（三分类）。使用不同的结果权重（二分类/三分类 = 5:5、2:8、1:9）和轮次（30、50、200）来选择最佳模型。通过准确率-损失曲线并计算曲线下面积（AUC）、准确率、敏感性、特异性、阳性预测值（PPV）和阴性预测值（NPV）来评估模型性能。
本研究共纳入了来自ASH图像库的115张骨髓涂片和来自临床的432张骨髓涂片。准确率-损失曲线结果表明，在结果权重和轮次为1:9和200的模型中观察到最佳的模型训练效果。同样，该模型在MDS的二分类以及AA、MDS、AML的三分类中也表现良好。测试集中MDS二分类模型的AUC、准确率和敏感性分别为0.985 [95%置信区间（CI），0.979 - 0.991]、0.914（95%CI，0.895 - 0.934）和0.992（95%CI，0.980 - 1.000）。测试集中AA、MDS、AML三分类模型的AUC、准确率和敏感性分别为0.968（95%CI，0.960 - 0.976）、0.929（95%CI，0.916 - 0.941）和0.857（95%CI，0.828 - 0.886）。
Image-net预训练模型能够获得较高准确率的AA、MDS、AML区分结果，可能为临床医生提供一种区分AA、MDS和AML的便捷工具。
