Bifurcated Auto Encoder based on Attention Mechanism

1 - Implementation of a network for segmentation of covid-19 infected regions in CT-images.
2 - A method based on generative models for generating synthetic infected regions.

For more details please refer to our paper

Different parts of the network are listed below :

• Encoder
• Decoder
• Inception Block
• Channel-wise Attention
• Spatial-wise Attention
• Attention Fusion

Architecture

The overall architecture of the network is shown below :

Inception Block

The modified inception block can be seen below :

Channel-Wise Attention

The modified version of this module is shown below :

Attention Fusion

In this block two attention modules are combined with a separate branch for residual connection :

Generating Synthetic Data

Data augmentation is an method for increasing the amount of data. In this work a data augmentation method based on generative models is proposed. A Pix2Pix conditional GAN is used as a network for converting binary infected regions to real infected regions.

Pix2Pix Conditional GAN

The overview of how the conditional GAN works can be seen here :

Overview

There are 3 main steps for segmenting the covid-19 infected regions :

• Augmentation
• Segmentation
• PostProcessing

Generating new data

For having some new data in the dataset we first trained a pix2pix GAN only on the infected regions of dataset. By doing so, the network has learned the process of converting a binary infected region to a real one. Then the infected part was replaced with the one in dataset.

References

• pix2pix tutorial