论文部分内容阅读
目的:开发一种基于3D-CNN的脑部多发性硬化症（MS）病灶图像的自动分割方法。方法:该分割方法分为两个阶段，包括2个CNN卷积层和池化层。第一阶段初步筛选出病灶体素，第二阶段进一步限定条件，从第一阶段得到的病灶体素中挑选出最终的分割结果。在MICCAI2016公共数据集上，对所提出的方法进行实验验证，并与其他的基线方法进行比较。结果:对于15位MS患者的MRI图像，所提出方法得到的平均相似性系数（DSC）为0.59，相比于3个基线方法，分别提高了2%、3%和4%。结论:所提出基于3D-CNN的图像分割方法在3D空间上对MRI图像进行分割，相比2D图像分割，对于临床诊断更具意义。“,”Objective:To develop an automatic segmentation method for multiple sclerosis (MS) brain lesions based on 3D-CNN.Method:The segmentation method was divided into two stages, including two CNN convolutional layers and a pooling layer. In the first stage, the voxels of the lesions were preliminarily screened out. In the second stage, the conditions were further limited, and the final segmentation results were selected from the voxels of the lesions obtained in the first stage. On the MICCAI2016 public data set, the proposed method was experimentally verified and compared with three baseline methods.Results:For MRI images of 15 MS patients, the average similarity coefficient (DSC) obtained by the proposed method was 0.59, which was an increase of 2%, 3%, and 4% compared with the three baseline methods, respectively.Conclusions:The proposed 3D-CNN-based image segmentation method segment MRI images in 3D space, which is more meaningful for clinical diagnosis than 2D image segmentation.