Deep Generative Models in the Real World: An Open Challenge from Medical Imaging - shubham223601/Anomaly-Detection GitHub Wiki

Referenced from http://arxiv.org/abs/1806.05452

This paper suggests to model the pixel wise probability using healthy data and detects abnormalities as pixel with low probability values. Approach is to estimate the distribution of healthy data Ph so as to evaluate the probability of each pixel of an unseen image. In this evaluation stage probability of abnormal regions would be low where as the probability of healthy regions would be higher.

various methods exist for estimating distribution of healthy data such as Autoencoder, consisting of encoder and decoder, where encoder maps the input to low dimensional latent variable and using this latent variable decoder reconstructs the input. Downside of AE is since it relies on lower dimensional representation for reconstruction, it is not able to reconstruct the variations which it has not seen during training time.

A better approach as compared with autoencoder is use of Denoising autoencoders, where the concept is similar as that of autoencoder however in case of denoising autoencoder, image is reconstructed back from the corrupted images(some noise is added to the actual image, so that it learns variation)

Variational AE and AAE are generative models, which model ph by taking into consideration of variational inference by instaead of encoding an input as a single point, input is encoded as distribution over latent space, then a point is samples from the latent space of that distribution which is decoded and the reconstruction error is computed. loss function is comprised of a reconstruction term + regularization term that tends to regularize the organisation of the latent space by making the distributions returned by the encoder close to a standard normal distribution in case of Variational AE where as in case of AAE, the distribution need not be necessarily normal.

VAE approach tend to generate blurred images which is resolved in case of AAE because of use of adverserial learning