一种基于注意力机制的CT图像预处理方法

Preprocessing Method for CT Image Based on Attention Mechanism

在CT图像分割的过程中,通常都会先使用窗口技术对图像数据做预处理.然而,由于脏器组织与病变组织密度的不均匀性,使得这种预处理方式并不能完全将无效信息除去而重点关注脏器和病变部位.受注意力机制的启发,本文提出了一种新的CT图像预处理方式.首先,根据专家经验制作标签,计算出组织的空间概率分布,利用概率分布制作出CT蒙版.其次,将蒙版覆盖到原图上后即可除去大量低概率区域的图像信息,降低网络学习难度.对于蒙版的制作,文中给出了3种做法.针对由于蒙版覆盖导致信息损失的特性,本文还提出了一种蒙版和初次分割结果相结合的方式,用于补偿使用蒙版导致的信息损失.本文提出的预处理方法在经典网络结构FCN、U-net和SegNet,在LiTS和3Dircadb数据集上做了多组对比,实验结果表明该方法对CT图像的分割结果有较好的效果.

Windowing technology is usually used to preprocess the image data during CT image segmentation.However,due to the inhomogeneity of the density between visceral tissue and lesion tissue,this method can not remove the invalid information completely and focus on the visceral and lesion.Inspired by the attention mechanism,this paper proposes a new CT image preprocessing method.First of all,the CT mask will be produced based on the probability distribution of the orgnization according to the labels from the experts′experience.Then,after multiplying the mask cover back to the original CT image,a large number of low probability area information will be removed,so as to reduce the difficulty of network learning.In this paper,the production process of mask is described in detail,and three production methods are proposed.According to the characteristics of information loss caused by mask coverage,this paper also proposes a method of combining masking with initial segmentation results to compensate for the information loss caused by using masking.The method has compared with FCN,U-net and Segnet on LiTS and 3dircadb datasets,and the results shows that it has a good effect on CT image segmentation.