Reading report: V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation

V-Net模型：全卷积神经网络在体素医学图像分割中的应用

**Abstract:**Convolutional Neural Networks (CNNs) have been recently employed to solve problems from both the computer vision and medical image analysis fields. Despite their popularity, most approaches are only able to process 2D images while most medical data used in clinical practice consists of 3D volumes. In this work we propose an approach to 3D image segmentation based on a volumetric, fully convolutional, neural network. Our CNN is trained end-to-end on MRI volumes depicting prostate, and learns to predict segmentation for the whole volume at once. We introduce a novel objective function, that we optimise during training, based on Dice coefficient. In this way we can deal with situations where there is a strong imbalance between the number of foreground and background voxels. To cope with the limited number of annotated volumes available for training, we augment the data applying random non-linear transformations and histogram matching. We show in our experimental evaluation that our approach achieves good performances on challenging test data while requiring only a fraction of the processing time needed by other previous methods.

Milletari, Navab, & Ahmadi - 第四届三维视觉国际会议（3DV）首次发布于2016年
Title: 提出了一种基于卷积神经网络的全连接方法用于体素医学图像分割_

Authors: Fausto Milletari, Nassir Navab, Seyed-Ahmad Ahmadi
Appeared in the proceedings of the conference: 2016 Fourth International Conference on 3D Vision (3DV)

Introduction and related work

Prostate segmentation represents a difficult problem for CNNs, as the volume of the prostate must be evaluated and its boundary must be precise. Early methods predominantly utilized pixel-wise image classification. However, these segmentations primarily focus on local context and are inefficient. However, such networks were still mainly applied to 2D images. In this paper, the researchers presented an innovative volumetric end-to-end fully convolutional neural network incorporating a unique objective function.

Method

In the contracting pathway, for each step, the researchers constructed as a series of convolutions employing 5×5×5 filters with stride 2; coupled with an element-wise summation and down-convolution. The application of 5×5×5 convolutions, supplemented with padding, increased the number of channels by multiplication. After passing through PReLU non-linearity, they subsequently performed an element-wise summation to learn a residual function. The size of this volume remained equivalent to its input. Following this, in the right portion, they restored the contracting volume by substituting down-convolution with up-convolution (2x2 de-convolution with stride 2). Additionally, they incorporated fine-grained feature forwarding akin to U-Net architecture. Finally, two output volumes were applied through softmax function analysis to classify voxels into foreground or background probabilities: “The network's two output volumes... are processed through a softmax layer to output probabilities for each voxel belonging to foreground or background.” To enhance training robustness, they introduced a novel objective function based on dice loss.

The team's future work will focus on dividing volumes containing multiple regions in other imaging techniques, such as ultrasound, and at higher resolutions by distributing the network across several GPUs.