目的 运用深度学习方法构建卷积神经网络(CNN)模型，实现中期染色体的自动分类，并评价模型染色体分类的准确率。 方法 收集2013年1月至2019年7月在宁波市妇女儿童医院出生缺陷防治实验室接受染色体病检查者的3 300份样本。共纳入3 300×46张染色体图片，其中70%作为训练集，30%作为测试集，用于构建CNN模型，随后应用952张中期染色体图片对建立的模型的染色体计数及"切割＋识别＋排列＋自动分析"的准确率进行评价。另取80张染色体图片资料，分别记录人和模型完成染色体分类的时间及准确率，综合评价模型的应用价值。 结果 CNN模型的分裂相计数准确率为61.81%，单条染色体的"切割＋识别＋排列＋自动分析"准确率为96.16%。CNN模型的分类耗时相对人工操作有了大幅度的提高，核型分析的准确率仅比遗传专科医师低3.58%。 结论 基于CNN构建的染色体分类模型具有较高的分类能力，能够减轻遗传医师在核型分析的过程中手动切割分类染色体的劳动强度，提高工作效率，因而具有较好的应用前景。 Objective To train a deep convolutional neural networks (CNN) using a labeled data set to classify the metaphase chromosomes and test its accuracy for chromosomal identification. Methods Three thousand and three hundred individuals undergoing surveillance for chromosomal disorders at the Laboratory for Comprehensive Prevention and Treatment of Birth Defects, Ningbo Maternal and Child Health Care Hospital from January 2013 to July 2019 were enrolled. A total of 3 300×46 chromosome images were included, of which 70% were used as the training set and 30% were used as the test set for the deep CNN. The accuracy of chromosome counting and "cutting + recognition + arrangement + automatic analysis" of the model were respectively evaluated. Another 80 images were collected to record the time and accuracy of chromosome classification by geneticists and the model, respectively, so as to assess the practical value of the model. Results The CNN model was used to count the chromosomes with an accuracy of 61.81%, and the "cutting + recognition + arrangement + automatic analysis" accuracy of the model was 96.16%. Compared with manual operation, the classification time of the CNN model has been greatly reduced, and its karyotyping accuracy was only 3.58% lower than that of geneticists. Conclusion The CNN model has a high performance for chromosome classification and can significantly reduce the work load involved with the segmentation and classification and improve the efficiency of chromosomal karyotyping, thereby has a broad application prospect.
Application of convolutional neural networks for the classification of metaphase chromosomes
Objective To train a deep convolutional neural networks (CNN) using a labeled data set to classify the metaphase chromosomes and test its accuracy for chromosomal identification. Methods Three thousand and three hundred individuals undergoing surveillance for chromosomal disorders at the Laboratory for Comprehensive Prevention and Treatment of Birth Defects, Ningbo Maternal and Child Health Care Hospital from January 2013 to July 2019 were enrolled. A total of 3 300×46 chromosome images were included, of which 70% were used as the training set and 30% were used as the test set for the deep CNN. The accuracy of chromosome counting and "cutting + recognition + arrangement + automatic analysis" of the model were respectively evaluated. Another 80 images were collected to record the time and accuracy of chromosome classification by geneticists and the model, respectively, so as to assess the practical value of the model. Results The CNN model was used to count the chromosomes with an accuracy of 61.81%, and the "cutting + recognition + arrangement + automatic analysis" accuracy of the model was 96.16%. Compared with manual operation, the classification time of the CNN model has been greatly reduced, and its karyotyping accuracy was only 3.58% lower than that of geneticists. Conclusion The CNN model has a high performance for chromosome classification and can significantly reduce the work load involved with the segmentation and classification and improve the efficiency of chromosomal karyotyping, thereby has a broad application prospect.
Deep learningChromosome classificationKaryotypingConvolutional neural network
万方数据
