/Mutli-cnn-covid19-detection

This is a python regeneration of a paper published at elsevier.

Primary LanguageJupyter Notebook

Mutli-cnn-covid19-detection

This repository is a python regeneration of a this paper which is a matlab implementation for computer aided covid 19 detection using multi-cnn and Bayesnet (Brnoulli classifier in my case) from X-ray images.

Dataset The dataset used in this paper was gathered from this open source repo and this kaggle page.

I combined these two datasets and made a balanced dataset of 563 covid19 and 563 of other illnesses, but due to lack of resource and time I only used 400 of images in each category.

Preprocessing

I converted all of the images to black and white and saved them in a seperated folder but


  • Feature Extraction

We used 2 pre trained CNNs for feature extraction and extracted 1280 features from MOBLENET and 2048 features from XCEPTION and replaced each image with a 1x3328 matrix.

  • Feature selection

First, in this step I used a correlation based feature selection. I used both correlation and pearson-r for removing most correlated features, but I could not since there was none.

Second, I used a forward stepwise feature selection which helped me select 15 most important features.

  • Classification

I used many algorithms and even tried to tune them but the best performing classifier was bernoulli classifier which gave me a 86 percent accuracy.