Paper Review:Automatic Liver Segmentation Using U-Net with Wasserstein GANs
A short summary of http://www.joig.org/uploadfile/2018/1226/20181226015625255.pdf
The following paper presents Automatic Liver Segmentation off CT images using Modified U-nets and Wasserstein GANs.
The proposed method comes with these major takeaways:
>GAN’s generator: The generator uses a modified Unet.Unets are widely used in biomedical segmentation tasks. Read more about Unets here.
The modification in Unets: Abolition of pooling layers and replacing them with BatchNormalization layers.The activation of the encoder of this Unet is leaky RELU but that of the decoder is the vanilla RELU.
Click here for a simple implementation of Unets in Pytorch
>Discriminator: The architecture of the Discriminator is the same as the architecture of the encoder in Generator.
Why did we use WGAN?
- Normal GANs have the problem of vanishing gradient, instable mode collapse & learning.
- WGAN uses Wasserstein distance. WGAN is a minimax game of Generator G and Discriminator D as follows:
min𝐺 max𝐷∈Ð 𝔼𝑥~ℙ𝑟 [𝐷(𝑥)]− 𝔼𝑧~ℙ𝑧 [𝐷(𝐺(𝑧))]
where ℙ𝑟 is real data distribution and ℙ𝑧 is generated data distribution. 𝑧 is random number based on uniform distribution. 𝐷 is a function of 1-Lipschitz. WGAN finds the Wasserstein distance of the input real image and the generated image. As a result, the learning of WGAN is more stable than that of GAN.
Dataset and Evaluation Metrics
3D-IRCADb (3D Image Reconstruction for Comparison of Algorithm Database) is a database includes several sets of anonymized medical images of patients and the manual segmentation of the various structures of interest performed by clinical experts. Access to dataset
The metric used is DICE . DICE is very similar to F1-score
𝐷𝑖𝑐𝑒 = |𝑇 ∩ 𝑃| (|𝑇| + |𝑃|)/2
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = |𝑇 ∩ 𝑃| |𝑃|
𝑅𝑒𝑐𝑎𝑙𝑙 = |𝑇 ∩ 𝑃| |𝑇|
Thank you for reading this!
~by Sayantan Das
