徐胜舟,程时宇.基于全卷积神经网络迁移学习的乳腺肿块图像分割[J].中南民族大学学报自然科学版,2019,(2):278-284
基于全卷积神经网络迁移学习的乳腺肿块图像分割
Breast mass image segmentation based on transfer learning of fully convolutional neural networks
  
DOI:10.12130/znmdzk.20190223
中文关键词: 乳腺肿块  分割  全卷积神经网络  迁移学习
英文关键词: breast mass  segmentation  fully convolutional neural network  transfer learning
基金项目:国家自然科学基金资助项目(61302192);中央高校基本科研业务费专项资金项目(CZY19011)
作者单位
徐胜舟,程时宇 中南民族大学 计算机科学学院, 武汉430074 
摘要点击次数: 133
全文下载次数: 129
中文摘要:
      针对乳腺X线摄片中肿块通常会被周围致密组织所掩盖,对比度低,且其形状不规则,肿块图像分割困难的问题,设计了一种基于全卷积神经网络迁移学习的乳腺肿块图像分割方法.该方法首先对乳腺肿块图像进行数据增强,然后利用迁移学习,对设计的全卷积网络模型载入参数并训练分割模型,最后在训练好的模型上对待分割图像进行处理.分割结果采用区域面积重叠率、Dice相似系数、Hausdorff距离等指标进行评价分析,在公开数据集的483幅图像上的实验结果表明:提出的方法的分割效果明显优于传统分割算法.
英文摘要:
      Breast mass segmentation exists difficulties since masses in mammograms may appear with irregular shapes, low contrast and share a similar intensity distribution with the surrounding breast structures. A method of breast mass image segmentation based on transfer learning of full convolutional neural network is proposed. The method firstly makes data augmentation for the breast mass image, then uses the transfer learning to load the parameters of the designed full convolution neural network model and trains the segmentation model. Finally, the image to be segmented is processed on the trained model. The segmentation results are evaluated by the area overlap ratio, Dice similarity coefficient, Hausdorff distance and other indicators. The experimental results on the 483 images of the public data set indicate that the segmentation results of the proposed method are significantly better than the traditional segmentation algorithms.
查看全文   查看/发表评论  下载PDF阅读器
关闭