Abstract
Prostate diseases are very common in men. Accurate segmentation of the prostate plays a significant role in further clinical treatment and diagnosis. There have been some methods that combine the segmentation network and generative adversarial network, using the adversarial training to boost the performance of segmentation network. However, the traditional adversarial training is unstable, which is hard to train. This attribute can easily lead to training failure. In this paper, we propose a segmentation network with self-attention adversarial training based on Wasserstein distance to tackle the problem. First, a segmentation network with residual connection and attention mechanism is deployed to generate the prostate segmentation prediction. Then, a self-attention discriminator network is added to the segmentation network to discriminate the prediction from ground truth. In the discriminator network, we replace the cross-entropy loss function with Wasserstein distance loss function which is better to measure the difference between distributions. The comparative experiments suggest our method is more stable than traditional adversarial training and achieves state-of-the-art performance.
Highlights
Prostate diseases are very common in men and are generally judged by their Magnetic Resonance (MR) images
To address the dependencies between pixels in segmentation task and stabilize the process of adversarial training, we propose a self-attention adversarial training strategy based on Wasserstein distance to combine the segmentation network with discriminator network
METHOD we describe the details of our proposed segmentation network with self-attention adversarial training, from the segmentation network to the discriminator network, and the concepts of residual connection, channel attention, position attention, Wasserstein distance loss
Summary
Prostate diseases (e.g., prostate cancer, prostatitis and enlarged prostate) are very common in men and are generally judged by their Magnetic Resonance (MR) images. Most of them are based on U-Net [17], which has a high performance in semantic segmentation These methods are insufficient to learn both local and global contextual relations between pixels because they train their networks just with a pixel-wise cross-entropy loss. VOLUME 7, 2019 the liver in a 3D manner Their network tried to optimize a cross-entropy loss together with an adversarial term that aims to distinguish between the ground truth and predicted segmentation map. To overcome the limitation that pixel-wise crossentropy loss function cannot capture the relationship between pixels, Yang et al [20] introduced the adversarial training approach to segment the liver CT image, where a deep convolutional network was first deployed to generate liver segmentation, and a discriminator network was utilized to improve the shape consistency between prediction and ground truth. I=1 where c denotes the number of classes, y and ydenote the ground truth and prediction, respectively
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
