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

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

**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 - 2016 Fourth International Conference on 3D Vision (3DV) - 2016
Title: V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation

Author: Fausto Milletari, Nassir Navab, Seyed-Ahmad Ahmadi
Published: 2016 Fourth International Conference on 3D Vision (3DV)

Introduction and related work

Prostate segmentation is a challenging task for CNNs, since the volume of the prostate needs to be assessed and its boundary needs to be accurate. Early approaches adopted patchwise image clasification. However, such segmentations only focus on local context and they are not efficient. Fully CNNs trained end-toend were still only applied to 2D images. In this paper, the researchers proposed a volumetric end-to-end fully convolutional neural network with a novel objective function.

Method

In the contracting path, for each steps, the researchers designed a combination of several convolutions with 5x5x5 filter, stride 2; an element-wise sum and down-convolution. By 5x5x5 convolutions with certain paddings, the researchers multiplied the number of channels. PReLU non-linearity was followed after the convolutions. Then they add up the volume element-wise and learned a residual function. The size of the volume now is the same as input. Lastly, the researcher commited “down convolution” (2x2 filters with stride of 2). After that, the consequent volume would be sent into the next step.
Similarly, the contracting volume was restored in the right portion. The down convolution was replaced with “up convoluton” (2x2 de-convolution with stride of 2).
Meanwhile, a fine-grained feature forwarding was attached like U-net.
Finally, two ouput volume was applied to softmax function to give foreground or background result:
“The network predictions, which consist of two volumes having the same resolution as the original input data, are processed through a soft-max layer which outputs the probability of each voxel to belong to foreground and to background. ”
The researchers proposed a novel objective function based on dice loss to avoid the problem that the learning gets trapped in local minima and tends to bias towards background.

The team’s future work will aim at segmenting volumes containing multiple regions in other modalities such as ultrasound and at higher resolutions by splitting the network over multiple GPUs.