见康华美最新研究成果——人工智能助力白血病的智能化诊断
慢性粒细胞白血病(CML)具有特征性的形态学改变,目前形态学检查仍依赖于病理医生在显微镜下评估,耗时长、易受主观因素影响。
近日,见康华美与南京信息工程大学合作在美国病理学杂志(The American journal of pathology)发表题为“The diagnosis of chronic myeloid leukemia with deep adversarial learning”的文章。本研究建立了一种新的骨髓细胞分割模型——CMLcGAN及自动化诊断系统(Figure 1),实现了形态学上对CML的自动化辅助诊断。
Figure 1. The framework of the automatic diagnosis of CML. A. The procedure of CMLcGAN training included end-to-end data set construction and model training; B. The pipeline of the whole diagnostic process. The extraction of clinical features was based on the segmentation results of the trained CMLcGAN. The trained generator generated the segmentation results, and the training of the discriminator was only to supervise the generator's training. Since the effective diagnostic regions of each WSI were different due to the production of slices, SSTD was used to eliminate the difference in statistical features extracted among various cases.
本团队开发的CMLcGAN模型,用于从骨髓活检图像中分割巨核细胞。选取517幅骨髓活检图像进行CMLcGAN的性能评估,其平均像素精度(PA)为95.1%,平均交并比(IoU)为71.2%,平均Dice系数(DICE)为81.8%。同时,将CMLcGAN与其他七种基于深度学习的分割模型进行比较(Table 1),结果表明,CMLcGAN的各项指标性能均优于其它分割模型。
Table 1. The comparison and evaluation results of CMLcGAN and other seven models.
在临床验证阶段,提取了骨髓细胞的七维数据特征(Table 2),采用t检验分析,最终选取具有统计学意义的五维数据特征作为临床预测特征集,用于区分CML及正常对照。
Table 2. The P values of the extracted features.
选取8种二分类器,应用筛选的五维数据特征进行临床预测,同时使用三折交叉验证,通过ROC曲线的AUC值评判分类效果。8种二分类器均表明提取的五维数据特征能够将CML和正常对照有效区分,可辅助诊断CML(Figure 7)。
Figure 7. The ROC curve of the cross-validation result. All 8 binary classifiers showed that the extracted statistical features had diagnostic capabilities
本研究成果进一步推动了AI在形态学分析上的应用。见康华美作为天津市血液病理智能化诊断企业重点实验室,在形态学、流式细胞学、分子遗传学等多个方面开展人工智能辅助诊断的研究,致力于研发智能化诊断系统,从而推动血液病诊断模式的变革。
撰稿:王哲
审核:蔺亚妮
参考文献:https://ajp.amjpathol.org/article/S0002-9440(22)00119-5/fulltext