Lung adenocarcinoma CT images classification based on multi-task contrastive self-supervised dual network
Objecctive The accurate diagnosis of lung adenocarcinoma based on CT images is of great clinical significance for subsequent treatment.Convolutional nerual network (CNN) image classification methods mostly focus on the local features of images,and it is difficult to fully capture global knowledge and spatial features.In order to fully learn these effective features,this paper proposes a multi-task contrastive self-supervised dual network (MTCSSDN) to realize computer-aided diagnosis of this disease.Methods Firstly,the conformer-based backbone network is used to integrate 3D CNN and Transformer into a network framework.The feature coupling unit (FCU) is used to interactively fuse the local features and global representations under different resolutions.Then,a concurrent structure is used in the same network framework for cross-architecture contrast learning to jointly capture the rich spatial features of images,so as to train the pre-trained model with stronger feature expression performance.Finally,it is transferred to the downstream image classification task,and the original image is used to fine-tune the downstream network model to further improve the classification performance of the model.Results MTCSSDN algorithm is evaluated on the lung adenocarcinoma dataset,and achieves the mean classification accuracy of 79.70％±2.13％,sensitivity of 78.70％±4.22％,specificity of 74.00％±7.44％ and Youden index of 52.70％±6.12％.Conclusions Therefore,The MTCSSDN algorithm proposed in this paper can effectively improve the performance of auxiliary diagnosis of lung adenocarcinoma and has potential clinical application value.
classification of lung adenocarcinomaself-supervised learningconvolutional neural networkTransformercontrastive learning
万方数据
维普
