计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (24): 234-241.DOI: 10.3778/j.issn.1002-8331.2105-0092

• 图形图像处理 • 上一篇    下一篇

改进Mask R-CNN网络在医学图像识别与分割中的应用

卢苇,刘丹,邵敏,吴扬东   

  1. 1.贵州大学 现代制造技术教育部重点实验室,贵阳 550025
    2.贵阳市口腔医院 口腔颌面外科,贵阳 550002
  • 出版日期:2021-12-15 发布日期:2021-12-13

Application of Improved Mask R-CNN Network in Medical Image Recognition and Segmentation

LU Wei, LIU Dan, SHAO Min, WU Yangdong   

  1. 1.Key Laboratory of Advanced Manufacturing Technology of Ministry of Education, Guizhou University, Guiyang 550025, China
    2.Department of Oral and Maxillofacial Surgery, Guiyang Stomatological Hospital, Guiyang 550002, China
  • Online:2021-12-15 Published:2021-12-13

摘要:

针对现有医学图像处理方法在人体复杂结构组织器官分割中的不足,提出复用低层特征信息的Mask R-CNN网络。该网络可对特定组织器官识别时同时进行分割,为了提高包含较多细节信息的低层特征层的利用率,将低层的特征信息添加到高层的特征中,使低层与高层特性优劣互补,将原始图像首次长宽压缩两次后的特征层定义为C1层,而后分别通过复用C1层和复用依次卷积的C1层这两种方法实现。并将主干网络进行了精简,以加快网络的训练速度,降低识别和分割的时间。以下颌骨作为应用对象,自建包含1?064张下颌骨CT图片的数据集,按9∶1的比例划分为训练集和验证集进行训练,使得复用依次卷积C1层的Mask R-CNN网络的训练损失降至2.8%,验证损失降至6.6%,表明该网络在下颌骨的识别和分割上具有很高的准确率。

关键词: 神经网络, 特征融合, 医学图像处理, 下颌骨识别与分割, 时间成本

Abstract:

Aiming at the deficiencies of the existing medical image processing methods in the segmentation of human body complex structures, tissues and organs. A Mask R-CNN network that reuses low-level feature information is proposed, which can segment specific organs at the same time. In order to improve the utilization of the lower-level feature layer that contains more detailed information, the lower-level feature information is added to the high-level features, so that the low-level and high-level features complement each other. The characteristic layer of the original image after the first length and width compression twice is defined as the C1 layer, and then it is implemented by two methods of multiplexing the C1 layer and multiplexing the successively convolved C1 layer. And the backbone network is streamlined to speed up the network training speed and reduce the recognition and segmentation time. The mandible is used as the application object. A self-built data set containing 1 064 mandibular CT images is divided into a training set and a validation set at a ratio of 9∶1 for training, and then the Mask R-CNN network that convolves the C1 layer in sequence is reused. The training loss is reduced to 2.8%, and the verification loss is reduced to 6.6%, it indicates that the network has a high accuracy in the recognition and segmentation of the mandible.

Key words: neural network, feature fusion, medical image processing, mandible recognition and segmentation, time costs