U-Net
Introduction
The U-Net is a type of convolutional neural network (CNN) that was developed for biomedical image segmentation at the Computer Science Department of the University of Freiburg, Germany. The network is based on the fully convolutional network and its architecture was modified and extended to work with fewer training images and to yield more precise segmentations. U-Net is named after its U-shaped architecture.
Architecture
The U-Net architecture is symmetric and consists of two paths: a contracting path (encoder) and an expanding path (decoder). The contracting path is a typical CNN that consists of repeated application of convolutions, ReLU activation function, and max pooling operations. Each step in the contracting path doubles the number of feature channels. The expanding path, on the other hand, consists of an up-convolution followed by a concatenation with the correspondingly cropped feature map from the contracting path, and two regular convolutions.
The U-Net combines the location information from the contracting path to finally obtain a general information combining localisation and context, which is necessary to predict a good segmentation map. Unlike the fully convolutional network, the U-Net predicts the segmentation map for the full image size, which eliminates the need for an explicit patch-based approach.
Training
Training of the U-Net is performed with a pixel-wise soft-max over the final feature map combined with the cross-entropy loss function. The network is trained to minimize the loss. During training, a large amount of labelled training data is required. However, the U-Net is capable of learning from a smaller dataset due to its powerful data augmentation which is achieved by applying elastic deformations to the available training images.
Applications
U-Net has been widely used in biomedical image segmentation, particularly for the segmentation of neuronal structures in electron microscopic stacks. It has also been successfully used in a variety of other applications such as lung segmentation in chest X-rays, liver segmentation in CT images, and segmentation of various structures in MRI images.
In addition to biomedical imaging, U-Net has also been used in other fields such as geosciences for oil reservoir segmentation, in remote sensing for building segmentation, and in autonomous driving for road segmentation.
Advantages and Limitations
The main advantage of U-Net is its ability to produce high-quality segmentation results with less training data. It achieves this through extensive data augmentation. U-Net also outperforms many other network architectures in terms of segmentation accuracy due to its unique architecture that combines the location information from the contracting path with the contextual information from the expanding path.
However, U-Net also has some limitations. The quality of the segmentation result is highly dependent on the size of the training dataset and the variation in the data. If the training data is not representative of the entire population, the network may not perform well. Additionally, U-Net is computationally intensive due to its deep architecture.