作者:梁義钦， 赵司琦， 王海涛， 何勇军
Authors:LIANG Yi-qin, ZHAO Si-qi, WANG Hai-tao, HE Yong-jun摘要:摘 要:异常细胞检测是宫颈癌智能辅助诊断的关键技术，直接影响着检测系统的性能。但宫颈异常细胞大多以簇团的形式存在，细胞相互粘连、复杂多样，给异常细胞检测带来了挑战。为解决这一问题，本文提出了一种两阶段簇团宫颈异常细胞检测方法。该方法在第一阶段采用YOLO-v5目标检测网络，利用可变形卷积替换网络中的标准卷积，使卷积核的大小和位置可以根据当前病理图像内容进行动态调整，以适应不同簇团宫颈细胞的形状、大小等几何形变。在第二阶段利用监督对比学习网络学习正、异常簇团宫颈细胞之间的特征差异，实现高准确率的正、异常簇团宫颈细胞分类。实验表明，簇团宫颈细胞召回率达到89.69%，相比基线网络YOLO-v5提升了1.43%，正、异常簇团宫颈细胞分类准确率达到87.81%，相比基线网络ResNet提升了10.31%。
Abstract:Abstract:Abnormal cell detection is a key technique for intelligent assisted diagnosis of cervical cancer, which directly affects the performance of the detection system. However, most cervical abnormal cells exist in the form of clusters. Cells adhere to each other, complex and diverse, which brings challenges to abnormal cell detection. To solve this problem, we proposed a two-stage detection method for cluster cervical abnormal cells. In the first stage, we use YOLO-v5 target detection network. The standard convolution in the network is replaced by deformable convolution. The size and location of convolution kernel can be dynamically adjusted according to the current pathological image content, so as to adapt to the shape, size and other geometric changes of cervical cells in different clusters. In the second stage, the supervised contrastive learning network is used to learn the feature differences between positive and abnormal clusters of cervical cells, so as to achieve high accuracy classification of positive and abnormal clusters of cervical cells. The experimental results show that the recall rate of cluster cervical cells reaches 89.69 %, which is 1.43 % higher than that of baseline network YOLO-v5.The classification accuracy of positive and abnormal cluster cervical cells reaches 87.81 %, which is 10.31 % higher than that of baseline network ResNet.

PDF全文下载地址:

可免费Download/下载PDF全文