Virtual Adversarial Training for Cross Domain Patch Contrastive Learning for Semi-supervised Cellular Nuclear Semantic Segmentation
To solve the problem that the semantic segmentation quality of semi-supervised contrastive learning is highly dependent on the prediction of smooth and correct pseudo labels,a semi-supervised nuclear semantic segmentation method based on cross-domain patch contrastive learning for virtual adversarial training is proposed.The proposed method integrates virtual adversarial training(VAT)into the cross-domain patch contrastive learning semi-supervised cellular nuclear semantic segmentation model to improve the smoothness and accuracy of the network prediction of pseudo label,and the consistency regularization loss of pixel self-weighting is used to replace the o-riginal consistency regularization loss of manually set high confidence threshold,and the loss of each pixel in the image is self-weighted for correct and effective use of pseudo label for network prediction.The experimental results show that at the ratio of 1/32,1/16 and 1/8 of the labeled images,on the MoNuSeg dataset,the Dice coefficient and Jaccard coefficient of the proposed method improved by 0.96 per centage points and1.11 per centage points,0.74 per centage points and0.85 per centage points,1.40 per centage points and 2.00 per centage points,respectively,compared with the CDCL model.On DSB dataset,Dice coefficient and Jaccard coefficient increased by 1.69 per centage points and2.27 per centage points,1.47 per centage points and2.19 per centage points,1.24 per centage points and1.77 per centage points,respectively,compared with CDCL model.
万方数据
维普
