Healthy, Pneumonia, and Covid-19 chest X-ray classification of V7 Labs & CloudFactory Release Annotated X-Ray Dataset

• Link: https://github.com/v7labs/covid-19-xray-dataset
• Blog: https://blog.cloudfactory.com/annotated-chest-x-ray-dataset-for-covid-19-research

I used the notebook NB.ipynb to transform the ground truth to work for my setting.

It consisted of ~6000 images with labels when I did this project: Healthy, Pneumonia, and COVID-19 In this work I have applied two classification strategies using popular U-Net architecture.

1. Multi-label classification:

In this setting the labels will be [h, p, c], where each value will be either 0 or 1. For example, if any image is healthy, the label would be [1, 0, 0]. If any image has only Pneumonia it will be [0, 1, 0]. In most cases if there is COVID-19 present, there will also be Pneumonia.

2. Binary Classification

• In this setting, we treat each image as either image with no-covid (label 0), or with covid(label 1) and do binary classification.

How to run?

• Download the sample datasets.zip from releases(1400 images out of ~6.5k).
• Extract in the project directory (inside covidxfactory, so it should be covidxfactory/datasets/v7_lab_cf/...)
• Install pytorch and torchvision from official website
• pip install easytorch
• For full datasets, please visit the original source links provided above and use the notebook NB.ipynb to download the complete ~6k images.

python main.py -ph train -b 32 -wm multi -nw 16 -ep 71 -nch 1 -pat 11
python main.py -ph train -b 32 -wm binary -nw 16 -ep 71 -nch 1 -pat 11

10-fold cross validation results for multi-label classification:

SPLIT	Precision	Recall	F1	Accuracy
SPLIT_0.json	0.7627	0.8333	0.7964	0.9644
SPLIT_5.json	0.8654	0.9783	0.9184	0.9876
SPLIT_8.json	0.8605	0.7708	0.8132	0.9737
SPLIT_4.json	0.7544	0.9556	0.8432	0.9752
SPLIT_3.json	0.7551	0.9024	0.8222	0.9752
SPLIT_1.json	0.7903	0.9074	0.8448	0.9721
SPLIT_9.json	0.86	0.86	0.86	0.9783
SPLIT_6.json	0.7458	0.9167	0.8225	0.9706
SPLIT_2.json	0.7719	0.9778	0.8627	0.9783
SPLIT_7.json	0.8113	0.86	0.8349	0.9737
Global	0.7948	0.894	0.8415	0.9749

10-fold cross validation results for binary classification:

SPLIT	Precision	Recall	F1	Accuracy
SSPLIT_0.json	0.7963	0.7963	0.7963	0.9659
SPLIT_5.json	0.8627	0.9565	0.9072	0.9861
SPLIT_8.json	0.8039	0.8542	0.8283	0.9737
SPLIT_4.json	0.7586	0.9778	0.8544	0.9768
SPLIT_3.json	0.8293	0.8293	0.8293	0.9783
SPLIT_1.json	0.8182	0.8333	0.8257	0.9706
SPLIT_9.json	0.746	0.94	0.8318	0.9705
SPLIT_6.json	0.7778	0.875	0.8235	0.9721
SPLIT_2.json	0.8222	0.8222	0.8222	0.9752
SPLIT_7.json	0.8214	0.92	0.8679	0.9783
Global	0.8011	0.8794	0.8384	0.9748

Training: Validation:

References for U-NET

Architecture used, O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in MICCAI, (2015)