基于对比学习的半监督肝脏血管分割方法OA北大核心CSTPCD

Aiming at the problem that existing methods often depend on high-quality labeled data for model training,with a tendency to focus excessively on local information while neglecting global information,a semi-supervised liver vessel segmentation method was proposed based on global and local contrast learning.First,the Mean Teacher model was used as a framework to globally contrast the high-dimensional features output by the encoder,the global consistency of the features was captured,and richer global context information was obtained.Then,local contrast on the decoder output was performed to obtain the local pixel level in semantic segmentation features.Finally,the distribution difference minimization method was introduced to reduce the distribution difference between labeled and unlabeled data using the discriminator for adversarial training,which improved the generalization performance of the model.Experiment results show that the proposed method achieves remarkable results in the liver vessel segmentation task,with Dice of 74.36%,Jaccard of 59.73%,average surface distance(ASD)of 2.65 mm,and 95%Hausdorff distance(95HD)of 13.57 mm on the 3Dircadb dataset,which is superior to other semi-supervised methods.