Team: Brave Metropolitans

• Máté Levente Kis
• Nándor Réfi
• Miklós Bendegúz Szócska

Project Video Report

YouTube link: https://youtu.be/-Irc5UeJnYg

Datasets used in the project:

• X-Ray Image Dataset: https://github.com/muhammedtalo/COVID-19/tree/master/X-Ray%20Image%20DataSet
• nabeelsajid917: https://www.kaggle.com/nabeelsajid917/covid-19-x-ray-10000-images
• chest_xray: https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia
• COVID-19 Dataset: https://data.mendeley.com/datasets/8h65ywd2jr/3

Current state of development:

Our goal is to detect efficiently non-infected, bacterial infected or COVID-19 infected lungs on chest xrays. We decided to use transfer learning to reach our goal.

We are working with 3 labels:

• no_findings: no detected infection
• covid_19: infected with COVID-19
• pneumonia: bacterial infection

Datasets used for learning and evaluation:

• train
  • covid_19(370)
  • no_findings(449)
  • pneumonia(450)
• test
  • covid_19(37)
  • no_findings(50)
  • pneumonia(50)

So far we used transfer learning on the following models:

Model	Branch	File	Validation accuracy	Test accuracy
VGG19 with ImageNet	mate_workspace_vgg19	COVIDDetector.ipynb	Validation accuracy: 87.3%	Test accuracy: 85.4%
ResNet34 with ImageNet	mate_workspace_resnet34	COVIDDetector.ipynb	Validation accuracy: 83%	Test accuracy: 83.2%
Inception v3 with ImageNet	mate_workspace	COVIDDetector.ipynb	Validation accuracy: 82.2%	Test accuracy: 81.8%
ResNet152 v2 with ImageNet	nandor_workspace	ResNet152V2_covid.ipynb	Validation accuracy: 75.9%	Test accuracy: 81.0%
SEResNet152 with ImageNet	mate_workspace_seresnet152	COVIDDetector.ipynb	Validation accuracy: 88.14%	Test accuracy: 83.94%
DenseNet201 with ImageNet	miki_workspace_densenet	COVIDDetector.ipynb	Validation accuracy: 86.95%	Test accuracy: 83.94%

How to use:

Open any notebook in Google Colab and run cells. The notebook will automatically download the prepared datasets from this repository.