U-net Matlab

[Lorin Searing]

Theu-net is convolutional network architecture for fast and precise segmentation of images. Up to now it has outperformed the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. It has won the Grand Challenge for Computer-Automated Detection of Caries in Bitewing Radiography at ISBI 2015, and it has won the Cell Tracking Challenge at ISBI 2015 on the two most challenging transmitted light microscopy categories (Phase contrast and DIC microscopy) by a large margin (See also our annoucement).

We provide the u-net for download in the following archive:u-net-release-2015-10-02.tar.gz (185MB). It contains the ready trained network, the source code, the matlabbinaries of the modified caffe network, all essential third partylibraries, the matlab-interface for overlap-tile segmentation and agreedy tracking algorithm used for our submission for the ISBI celltracking challenge 2015. Everything is compiled and tested only onUbuntu Linux 14.04 and Matlab 2014b (x64)

If you do not have a CUDA-capable GPU or your GPU is smaller thanmine, edit segmentAndTrack.sh accordingly (see there fordocumentation). If you have any questions, you may contact me atron...@informatik.uni-freiburg.de, but be aware that I can notprovide any support.

In recent times, whenever we wish to perform image segmentation in machine learning, the first model we think of is the U-Net. It has been revolutionary in performance improvement compared to previous state-of-the-art methods. U-Net architecture for image segmentation is an encoder-decoder convolutional neural network with extensive medical imaging, autonomous driving, and satellite imaging applications.However, understanding how the U-Net performs segmentation is important, as all novel architectures post-U-Net develop on the same intuition. We will be diving in to understand how the u net segmentation performs image segmentation. To enhance our understanding, we will also apply the U-Net for the task of brain image segmentation.

Image classification for two classes involves predicting whether the image belongs to class A or B. The predicted label is assigned to the entire image. Classification is helpful when we want to see what class is in the image.

Olaf Ronneberger and his team developed u net segmentation in 2015 for their work on biomedical images. It won the ISBI challenge by outperforming the sliding window technique by using fewer images and data augmentation to increase the model performance.

Sliding window architecture performs localization tasks well on any given training dataset. This architecture creates a local patch for each pixel, generating separate class labels for each pixel. However, this architecture has two main drawbacks: firstly, it generates a lot of overall redundancy due to overlapping patches. Secondly, the training procedure was slow, taking a lot of time and resources. These reasons made the architecture not feasible for various tasks. U-Net architecture for image segmentation overcomes these two drawbacks.

This encoder network consists of 4 encoder blocks. Each block contains two convolutional layers with a kernel size of 3*3 and valid padding, followed by a Relu activation function. This is inputted to a max pooling layer with a kernel size of 2*2. With the max pooling layer, we have halved the spatial dimensions learned, thereby reducing the computation cost of training the model.

In between the encoder and decoder network, we have the bottleneck layer. This is the bottommost layer, as we can see in the model above. It consists of 2 convolutional layers followed by Relu. The output of the bottleneck is the final feature map representation.

Now, what makes U-Net so good at image segmentation is skip connections and decoder networks. What we have done till now is similar to any CNN. The skip connections and decoder network separates the u net architecture from other CNNs.

A 1*1 convolution follows the last decoder block with sigmoid activation which gives the output of a segmentation mask containing pixel-wise classification. This way, it could be said that the contracting path passes across information to the expansive path. And thus, we can capture both the feature information and localization with the help of a U-Net.

We use a publicly available dataset. This brain tumor T1-Lighted CE-MRI image dataset consists of 3064 images. There are 1047 coronal images, 990 axial images, and 1027 saggital images. This dataset has a label for each image, identifying the type of tumor. These 3064 images belong to 233 patients. The dataset includes three types of tumors- 708 Meningiomas, 1426 Gliomas, and 930 Pituitary tumors, which are publicly available on: Click Here.

We download the dataset from the link given above. The dataset needs to be unzipped and made available. Our dataset is given in matlab format, so we convert this data into numpy arrays each for the images, labels, and masks. We finally display the size of each numpy array.

Next, we normalize the input images. This step is followed by defining our evaluation metrics. We use binary cross-entropy, dice loss, and a custom loss function composed of these two. We also use a 80:10:10 ratio for our train:val:test split and display the size of each.

3a8082e126