Liver Segmentation in CT Images Using Deep Neural Networks

Automatically extracting the liver from CT or MR images due to its heterogeneous shape and proximity to other organs is a challenging task. In recent times, Deep Learning have shown good results in medical image segmentation. Among the developed networks, U-Net has recorded many successes in medical image segmentation. This research presents an algorithm to perform a detailed liver segmentation. In this algorithm, images are first classified with a classification network to be separated into the liver included and non-liver included classes, then the class containing the liver are analyzed with the segmentation network. The segmentation network is an extended version of the U-Net, which takes full advantage of ConvLSTM, densely convolutional layers, recurrent and residual blocks. In the construction and extraction path, common convolutional blocks have been replaced by R2Conv blocks, to train the network more abstractions from input features and prevent gradient vanishing. Also, the mechanism of densely convolutional layers has been used in the last convolutional layer of the construction path. This idea improves the power of network representation by allowing information propagation through the network and reusing features. To concatenate the feature maps in the corresponding contracting path and the up-sampled output, instead of a simple concatenation in skip connections, ConvLSTM was used. Finally, applying this algorithm to the data used in the trial CHAOS challenge for CT, has resulted in a Dice value of %97.5.