Aiming at these limitations of traditional lung segmentation methods which manual feature extraction and prior knowledge are needed,a convolutional neural network(CNN) based auto-matically segment method for lung fields in chest X-ray (CXR) is proposed,original image is splited into the left and right lungs and cut blocks from them as the training samples,and then use deep CNN to automatically discovery the potential characteristics, and then classify the image blocks. Secondly the classified results are mapped into two-value images as the initial segmentation results. Lastly the initial results are merged to get the area of lung fields. Our experimental results on the public JSRT dataset show that Jaccard metric of our proposed method can reach 94. 6%,mean boundary distance(MBD) metric can reach 1. 10 mm,and outperform current reported algorithms.%针对传统胸片肺野分割方法需要人工干预、提取特征以及对先验知识的依赖性问题,提出了一种基于卷积神经网络(CNN)的胸片肺野自动分割方法,将X光胸片的分割问题转换为图像块的分类问题.将原图像分割成左、右肺,切块处理后分别作为训练样本,利用深度学习自动发现图像块中的潜在特征,对图像块进行分类,并将结果映射成二值图,得到初步分割结果,再对其进行后处理,合并之后作为最终的分割结果.实验表明:此方法在公开的JSRT数据集上进行测试,Jaccard指标可达94.6%,平均边界距离(MBD)指标达到1.10 mm,较现存分割算法更加出色.
展开▼
