Covid-X-ray

A project of a Computer Vision Deep Learning model for Covid-19 and pneumonias Xray images classification.

IMPORTANT!

All the conclusions from this project should undergo a thorough clinical study, so it is not possible to validate any conclusion without a valid clinical test.

The Project

This project has been developed in May 2020 and intends to evaluate a deep learning model for classification of Xray images coming from various kind of pneumonia included Covid-19, using keras and Tensorflow

Fig. 1: On the left Covid-19 XRay, on the right normal lungs Xray

Is there a way for a deep learning computer vision model to tell the difference between various king of pneumonia by Xray images?

How this repository is structured

Here is a description of folders and files in this repository

• dataset_for_classification - here you will find x-ray images split in 4 self explaining subfolders. The Covid Xray images are taken from the Covid project of Dr. Joseph Paul Cohen. Some randomly selected images have been taken only if they were covid positive and in PA (posteroanterior) view. The other images have been randomly taken from Kaggle Chest X-Ray Images Pneumonia dataset
• Scripts and Notebooks - a folder with scripts and Colab notebooks with code used for the pre-processing tasks and model training. So far i included the covid19_multiclass.py file, that i used to train the model in the case of a 4 classes scenario, and the covid_binary.py file that i used in the case of covid/no-covid scenario (everything is explained in detail in the project report).
• Plots - Plots with losses and accuracies in different scenarios. for every case Classification Reports are included and JSON files with history of the model training too. Note that in this case we achived 0.98 F-score for covid detection.

Fig. 2: Classification Report for binary Classification scenario

In the file Covid-project-report.pdf you will find the report and all the steps i followed to train the model and achieve some conclusions.

Conclusions and further investigations

• The trained model seems to be accurate on this dataset but what is really detecting? Remember that the Xray images come from 2 different datasets for what concerns Covid and No‐Covid, is there any chance the model is detecting this difference maybe in the way the X‐ray images were taken?
• The dataset is still too small for this purpose it would be useful to make the same training on a larger dataset