基于半监督多头网络的腰椎CT图像分割

Lumbar CT Image Segmentation Based on Semi-supervised Multi-head Network

针对医学图像分割任务中医学数据标注困难以及CT图像强度不均匀问题,提出一种基于半监督的多头分割网络SSMH-Net。SSMH-Net网络采用教师—学生训练架构,基于相同的分割模型V-Net,通过指数移动平均算法完成教师与学生模型的交互训练;采用Multi-Head方法估计模型预测的不确定性信息,指导分割模型在更可靠的目标中学习。在CTspine分割数据集上,SSMH-Net网络平均分割Dice系数达到95.70%,表现出较为优异的分割性能。

Aiming at the difficulty of medical data labeling in the tasks of medical image segmentation and the uneven intensity in CT images,semi-supervised multi-head segmentation network SSMH-Net was proposed.Teacher-student training architecture was adopted,based on the same segmentation model V-Net,and completed the interactive training of teacher-student models through the exponential moving average algorithm.Multi-Head method was adopted to estimate the uncertainty information of model predictions,which guided the segmentation model to learn from more reliable targets.On the CTspine segmentation dataset,the SSMH-Net network shows excellent segmentation performance,which has an average segmentation Dice coefficient of 95.70%.