Overview of context splicing network.
Overview of CS module.
Automated volumetric medical image segmentation algorithms are highly demanded in clinical practice. With the advent of fully convolutional network (FCN) in semantic segmentation task, models based on FCN are proposed for medical image segmentation, and U-Net is one of the most successful model. However, U-Net and variations of U-Net always sacrifice feature resolution to pursue high-level features. Additionally, capacity of medical image segmentation models to capture multi-scale features and generate long-range dependencies are rarely concerned. In this paper, we exploit the latest image processing methods and propose a context splicing network (CS-Net) to extract rich contextual information while preserve spatial information. CS-Net consists of three major components: feature encoder module, context splicing module (CS module) and feature decoder module. The encoder is a pre-trained ResNet101 with atrous convolution. The CS module splice atrous spatial pyramid pooling (ASPP) block with a self-attention block that incorporates multi-scale features and global context. The proposed method is evaluated on three public datasets: CHAOS, LiTS and KiTS, which are collected by computed tomography. Comprehensive results show that CS-Net outperforms the original U-Net and the recent proposed CE-Net for liver segmentation and kidney segmentation and has competitive tumor segmentation results.Some example results are showed below:
Sample results of liver segmentation on CHAOS.
Sample results of liver segmentation and tumor segmentation on LiTS
Sample results of kidney segmentation and tumor segmentation on KiTS.
In this repository, we provide the implementation of CS-Net which is based on nnUNet. It’s worth noting that we change the input from one channel to three channels.
You can search "dawn changed here" globally in the project to see where we have modified.
And we change the model from the original nnUNet model to CS-Net in a rude way, you can see generic_UNet.py for details.
To set up CS-Net, follow the exactly same steps nnUNet given. Search "set your dir" to find where to set your directory quickly.
This project is created based on nnUnet, deeplabv3+ and DANet. And the authors retain all the copyright of the related codes.