Research On Medical Image Organ Segmentation Based On Convolution Neural Network And Transformer

With the rapid development of artificial intelligence technology and medical imaging,medical image processing methods based on artificial intelligence methods have played an important role in the clinical field,and it has become a common way to process various images in the field of medicine by using deep learning methods.In this paper,an end-to-end automatic segmentation method for medical images is implemented combining various deep learning methods,and the main research content in this paper is as follows:(1)To tackle the problem that the traditional convolutional neural network(CNN)is inadequate in extracting global information of imagesl,we proposes a multi-organ segmentation model based on Transformer and CNN dual encoder network,which adequately combines the advantages of CNN and Transformer.The proposed model extracts global information and captures local information at the same time.The fusion module is also introduced in the segmentation model to fuse the convolutional features and Transformer’s feature sequences,and the feature fusion effectively smooth the decision boundary by using skip connection for cascading.Experiments show that the segmentation model achieves 80.68% in Dice score for multi-organ segmentation of the abdomen and 91.12% in Dice score for multi-organ segmentation of the heart image,which improves the multi-organ segmentation effect.(2)To address the lack problem of available labeled image data,a semi-supervised medical image organ segmentation model based on a two-student model is proposed in this paper.The weights of two-student semi-supervised model uses independent initialization,which solves the problem of close coupling between the teacher and student models.In order to better learn multi-layer features,multi-resolution feature fusion is introduced in the decoding module of the semi-supervised segmentation model.The experimental results show that the multi-resolution semi-supervised segmentation algorithm achieves 86.85% and 90.22% in terms of average dice similarity coefficients,and 77.26% and 82.29% in terms of average jaccard similarity coefficients by using only 10% and20% of labeled data trained on 3D left atrial data,respectively.The proposed method is better than the segmentation performance of traditional segmentation methods.