Figure 1
From: Unsupervised abnormality detection in neonatal MRI brain scans using deep learning
Autoencoder (AE) network employs an encoder-decoder structure for learning. Encoders sequentially abstract or compress (down-sample, \(\downarrow \times 2\)) the input data, G to find the most effective feature maps while decoders sequentially amplify or expand (up-sample \(\uparrow \times 2\)) the resulting feature maps to reconstruct output data, \({\hat{G}}\) that minimizes the error between G and \({\hat{G}}\), \(E = G - {\hat{G}}\). Each block within the AE consists of multiple convolutional neural network layers, rectifiers, and normalization layers that are tuned during learning using MRI data from seemingly normal populations.
