作者:盖晋平，秦健，何勇军，彭晨辉
Authors:GAI Jin-ping,QIN Jian,HE Yong-jun,PENG Chen-hui摘要:摘要:深度学习的发展有效地提升了宫颈细胞分类的准确率。深度神经网络的训练需要大量的标注数据。而宫颈细胞图像的标注需要专业的医生，且标注工作量大，成本高。这使得宫颈细胞图像标注数量不足，从而限制了宫颈细胞分类性能的进一步提高。针对以上问题，提出了一种有效利用临床中大量未标注数据的宫颈细胞分类方法。该方法首先采用SimCLR训练一个改进的ResNet网络对细胞进行特征提取。然后用全连接神经网络根据提取的特征信息进行分类预测。在宫颈细胞分类的实验中，该方法使用512张标注图像得到87.85%的准确率和77.10%的精确度，相比于对比方法更加优越。
Abstract:Abstract:The development of deep learning has effectively improved the accuracy of cervical cell classification. The training of deep neural networks requires a large amount of labeled data. However, the labeling of cervical cell images requires specialized physicians and the labeling workload is heavy and costly. This results in an insufficient number of cervical cell image annotations, thus limiting further cervical cell classification performance improvements. In response to the above problems, a cervical cell classification method that effectively utilizes the large amount of unlabeled data in the clinic is proposed. The method first uses SimCLR to train a modified ResNet network for the feature extraction of cells. Then a fully connected neural net is used to make classification predictions based on the extracted feature information. In the experiment of cervical cell classification, the method in this paper obtained 87.85% accuracy and 77.10% precision using 512 annotated images, which is more superior compared to the comparison method.

PDF全文下载地址:

可免费Download/下载PDF全文