Segmentation of Bone Structure in X-ray Images using Convolutional Neural Network
Segmentation of Bone Structure in X-ray Images using Convolutional Neural Network
复制标题
DOI：
10.4316/aece.2013.01015
复制
发表时间：
2013-01-01
影响因子：
0.8
通讯作者：
Holban, Stefan
中科院分区：
文献类型：
--
作者：
Cernazanu-Glavan, Cosmin;Holban, Stefan
分割过程是任何从图像中提取信息的自动方法所必需的第一步。对于 X 射线图像，通过分割，我们可以将骨组织与图像的其余部分区分开来。现在有多种分割技术，但总的来说，它们都需要人工干预分割过程。因此，本文提出了一种使用卷积神经网络 (CNN) 的 X 射线图像分割新方法。目前，卷积网络是图像分割的最佳技术。它们在包括医学领域在内的所有领域的广泛使用证明了这一事实。由于X射线图像尺寸较大，为了减少训练时间，本文提出的方法仅从整个图像中选择某些区域（最大感兴趣区域）。神经网络用作像素分类器，从而根据方形区域中的原始像素值产生每个像素（骨骼或非骨骼）的标签。我们还将介绍选择网络最终配置的方法，并与其他 3 种 CNN 配置进行比较分析。我们选择的网络在所有使用的评估指标（即扭曲误差、兰特误差和像素误差）上都获得了最佳结果。
The segmentation process represents a first step necessary for any automatic method of extracting information from an image. In the case of X-ray images, through segmentation we can differentiate the bone tissue from the rest of the image. There are nowadays several segmentation techniques, but in general, they all require the human intervention in the segmentation process. Consequently, this article proposes a new segmentation method for the X-ray images using a Convolutional Neural Network (CNN). In present, the convolutional networks are the best techniques for image segmentation. This fact is demonstrated by their wide usage in all the fields, including the medical one. As the X-ray images have large dimensions, for reducing the training time, the method proposed by the present article selects only certain areas (maximum interest areas) from the entire image. The neural network is used as pixel classifier thus causing the label of each pixel (bone or none-bone) from a raw pixel values in a square area. We will also present the method through which the network final configuration was chosen and we will make a comparative analysis with other 3 CNN configurations. The network chosen by us obtained the best results for all the evaluation metrics used, i.e. warping error, rand error and pixel error.