Skip to content

Latest commit

 

History

History
58 lines (39 loc) · 2.6 KB

README.md

File metadata and controls

58 lines (39 loc) · 2.6 KB

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 : Alt text

Inception Block

The modified inception block can be seen below :
Alt text

Channel-Wise Attention

The modified version of this module is shown below :
Alt text

Attention Fusion

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

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 :
Alt text

Overview

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

  • Augmentation
  • Segmentation
  • PostProcessing

Alt text

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. Alt text

References