虚拟对抗训练的跨域块对比半监督细胞核分割

Virtual Adversarial Training for Cross Domain Patch Contrastive Learning for Semi-supervised Cellular Nuclear Semantic Segmentation

针对目前细胞核的半监督对比学习语义分割质量高度依赖于平滑、正确伪标签的预测等问题,提出了虚拟对抗训练的跨域块对比学习半监督细胞核语义分割方法。该方法将虚拟对抗训练(VAT)方法融入到跨域块对比学习半监督细胞核语义分割模型中用以提升网络预测伪标签的平滑度与准确度,并使用像素自加权的一致性正则化损失替换原有的人工设置高置信度阈值的一致性正则化损失,对图像中各像素的损失自加权,正确地对网络预测的伪标签进行有效利用。实验结果表明,在有标签图片比例为1/32,1/16和1/8下,该方法在MoNuSeg数据集上的医学图像分割评估指标Dice系数和Jaccard系数分别较CDCL模型提升了0.96百分点和1.11百分点,0.74百分点和0.85百分点,1.40百分点和2.00百分点,在DSB数据集上的Dice系数和Jaccard系数分别较CDCL模型提升了1.69百分点和2.27百分点,1.47百分点和2.19百分点,1.24百分点和1.77百分点。

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 original 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 and 1.11 per centage points,0.74 per centage points and 0.85 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 and 2.27 per centage points,1.47 per centage points and 2.19 per centage points,1.24 per centage points and 1.77 per centage points,respectively,compared with CDCL model.