融合片内语义和片间结构特征的自监督CT图像分类方法  

A Self-Supervised CT Image Classification Method Incorporating Intra-Slice Semantic and Inter-Slice Structural Features

作　　者：曹春萍[1] 许志华 CAO Chunping;XU Zhihua(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《电子科技》2024年第7期43-52,共10页Electronic Science and Technology

基　　金：浙江省卫生健康委员会面上项目(2022KY122);浙江省中医药科技计划(2019ZA023)。

摘　　要：针对CT(Computed Tomography)图像分析存在人工标签稀缺、分类性能不佳等问题,文中提出一种融合片内语义和片间结构特征的自监督CT图像分类方法。该方法利用CT图像的层次结构特性和局部组成要素的语义特点,通过混淆切片生成算法对无标签的病灶部位图像进行处理,生成空间指数和混淆切片作为监督信息。在自监督辅助任务中利用ResNet50网络从混淆切片中同时提取与病灶部位相关的CT片内语义和片间结构特征,将学习到的特征迁移到后续医学分类任务中,使得最终模型从无标签数据中获得增益。实验结果表明,当被使用的有标签数据有限时,相比其他针对CT图像的二维模型和三维模型,所提方法的分类性能和标签利用效率更优。In view of the scarcity of artificial labels and poor classification performance in CT(Computed Tomography)image analysis,a self-supervised CT image classification method combining in-slice semantic and interslice structural features is proposed in this study.In this method,the hierarchical structure of CT images and the semantic features of local components are utilized to process the unlabeled lesion images through the confusion section generation algorithm,and the spatial index and confusion section are generated as supervisory information.In the self-supervised auxiliary task,the ResNet50 network was used to extract both the intraslice semantic and interslice structural features related to the lesion site from the confused sections,and the learned features were transferred to the subsequent medical classification task,so that the final model gained from the unlabeled data.The experimental results show that compared with other 2D and 3D models for CT images,the proposed method can achieve better classification performance and label utilization efficiency when the used labeled data is limited.

关 键 词：医学图像分类 三维医学图像处理 CT图像 自监督学习 迁移学习 小样本学习 片内语义特征 片间结构特征 ResNet50 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]