School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai, 200093, China.
School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai, 200093, China.
Tissue Cell. 2022 Feb;74:101677. doi: 10.1016/j.tice.2021.101677. Epub 2021 Nov 20.
Cervical cancer is the second biggest killer of female cancer, second only to breast cancer. The cure rate of precancerous lesions found early is relatively high. Therefore, cervical cell classification has very important clinical value in the early screening of cervical cancer. This paper proposes a convolutional neural network (L-PCNN) that integrates global context information and attention mechanism to classify cervical cells. The cell image is sent to the improved ResNet-50 backbone network to extract deep learning features. In order to better extract deep features, each convolution block introduces a convolution block attention mechanism to guide the network to focus on the cell area. Then, the end of the backbone network adds a pyramid pooling layer and a long short-term memory module (LSTM) to aggregate image features in different regions. The low-level features and high-level features are integrated, so that the whole network can learn more regional detail features, and solve the problem of network gradient disappearance. The experiment is conducted on the SIPaKMeD public data set. The experimental results show that the accuracy of the proposed l-PCNN in cervical cell accuracy is 98.89 %, the sensitivity is 99.9 %, the specificity is 99.8 % and the F-measure is 99.89 %, which is better than most cervical cell classification models, which proves the effectiveness of the model.
宫颈癌是女性癌症的第二大杀手，仅次于乳腺癌。早期发现的癌前病变治愈率相对较高。因此，宫颈细胞分类在宫颈癌的早期筛查中具有非常重要的临床价值。本文提出了一种卷积神经网络（L-PCNN），该网络集成了全局上下文信息和注意力机制，用于对宫颈细胞进行分类。将细胞图像发送到改进的 ResNet-50 骨干网络中，以提取深度学习特征。为了更好地提取深度特征，每个卷积块引入卷积块注意力机制，引导网络关注细胞区域。然后，骨干网络的末端添加一个金字塔池化层和一个长短时记忆模块（LSTM），以聚合不同区域的图像特征。整合低层次特征和高层次特征，使整个网络能够学习更多的区域细节特征，解决网络梯度消失的问题。实验在 SIPaKMeD 公共数据集上进行。实验结果表明，所提出的 l-PCNN 在宫颈细胞准确率方面的准确率为 98.89%，灵敏度为 99.9%，特异性为 99.8%，F1 测度为 99.89%，优于大多数宫颈细胞分类模型，证明了该模型的有效性。
