SC-SENet：一种基于阴道镜图像的宫颈上皮内瘤变分类诊断模型

SC-SENet: A Classification Diagnosis Model for Cervical Intraepithelial Neoplasia Based on Colposcopic Images

摘要: 宫颈上皮内瘤变(Cervical Intraepithelial Neoplasia, CIN)是宫颈浸润癌的癌前病变阶段。为了提高诊断的效率与准确性，本研究使用深度学习技术对阴道镜图像进行三分类识别，搭建了一个结合了Swin Transformer与卷积神经网络(CNN)的分类诊断模型Swin-Conv Squeeze-and-Excitation Network (SC-SENet)。该模型包括Swin Transformer分支与结合了通道注意力机制的卷积分支，通过并联的方式结合分别用于图像全局特征与局部特征的提取，结合两种特征实现病灶区域信息的特征识别以提高诊断准确度。本研究使用临床采集的阴道镜图像作为原始数据集，SC-SENet模型在CIN三分类诊断任务的总体准确率为90.81%。实验结果表明，本研究提出SC-SENet模型能有效诊断宫颈上皮内瘤变的病变等级。

Abstract: Cervical Intraepithelial Neoplasia (CIN) is a precancerous stage of invasive cervical cancer. In order to improve the efficiency and accuracy of diagnosis, this study used deep learning technology to classify colposcopy images into three categories. In this paper, a classification diagnosis model Swin-Conv Squeeze-and-Excitation Network (SC-SENet) combining Swin Transformer and Convolutional neural Network (CNN) is built. The model includes the Swin Transformer branch and the convolution branch combined with the channel attention mechanism, which are combined in parallel to extract the global features and local features of the image respectively. The two features are combined to realize the feature recognition of the lesion area information to improve the diagnostic accuracy. In this study, clinically collected colposcopy images were used as the original data set, and the overall accuracy of SC-SENet model in CIN three-classification diagnosis task was 90.81%. The experimental results show that the SC-SENet model proposed in this study can effectively diagnose the lesion grade of cervical intraepithelial neoplasia.