Deep Autoencoding Models for Unsupervised Anomaly Segmentation in Brain MR Images - shubham223601/Anomaly-Detection GitHub Wiki

Referenced from http://arxiv.org/abs/1804.04488

Approach utilizes spatial auto encoding model(AnoVAEGAN, consisting of fully convolutional encoder decoder architecture, which is used to preserve the spatial information in the latent space) instead of a patch based model to capture the variability of MRI image.

Main difference between traditional Auto encoders and AnoVAEGAN is AE are able to create high resolution images but they create blurry images because of memorization problems where as AnoVAEGAN builds model in form of spatial variational autoencoders to build a model that captures global appearence instead of local patches. The reconstruction objective then allows to train a generative model on complex and high resolution data

A form of VAEGAN is used to map the input image to a lower dimensional representation and back to high quality image reconstruction using the encoder and decoder. Here latent space Z Space is constrained to follow Multivariate normal distribution. Input is basically trained with the help of encode and decoder network capturing the spatial information, where decoder is used to reconstruct the image and the discriminator is used to discriminate between reconstructed and the original image. At encoder side mapping from input image to a lower dimensional representation z takes place and back to high quality image using decoder.

Model is optimized using two loss function in an alternating manner, one of which is for VAE and the other one for the discriminator

Here Loss function for VAE comprises of there terms :

  • Reconstruction loss - pixel wise l1 distance between orginal and reconstructed image
  • Prior loss - KL Divergence between MVN distribuion and the distribution of z, This loss acts as a regularizer
  • Adversial loss - which forces decoder to generate the image which is likely to fool discriminator in task of differentiating between actual and reconstructed image
⚠️ **GitHub.com Fallback** ⚠️