Covid 19 detection from chest x ray using Deep Learning

As part of my MS Computer Science Machine Learning course project, we were asked to read a research paper and then implement that paper. I was more interested in papers related to Covid 19 data. So, I came across with a paper which proposed the idea of predicting Covid 19 from the chest X ray. They consult with medical board which help them in finding some abnormalities on covid19 patients which gave them confidence to build a model that can predict Covid 19 from chest x-ray images.

Data Collection:

Publicly available data was used by the authors. They got 250 samples of Covid-19 chest x-ray and they get 5000 samples of other dieses from publicly available ChesPert dataset. As mention earlier they consult with medical team over the Covid-19 patient chest x ray’s, they remove around 66 chest x-rays because image was not clear. Training set contain 2000 non Covid images and 84 covid images. Covid 19 images sample is very low so data augmentation was applied to increase training sample to 420 on Covid19 images. Testing was done with 100 Covid 19 images and 3000 non Covid.

Classification Methods

Deep transfer learning was used for the predictions. Res18, Res50, SequeezeNet and DenseNet models were trained for the prediction. These models are based Convolution Neural Networks approach which is very much common for the image processing and feature extractions.

Pre trained model were used as we had a very limited data sample available for Covid=19. In which model parameter were fine tuned with new data.

Classification were made on 2 classes one “Covid” and other “Non Covid”.

Models were implemented using Pythorch library that is publicly available.

Results:

Results were good we got accuracy of 98% for all the models. Sensitivity and specificity rates were also promising. Models were evaluated using different parameters like ROC, probability, confusion matrices.

Limitation:

Data available for Covid-19 is very low that’s why we cannot reliable on results.

My contribution:

In adding to the reproduction of work done in research paper, I further trained model on VGG, Inception V3 and alexNet. These are also very good CNN based models and they gave accuracy up to 99%.

Further more I found new dataset with 1200 samples of Covid Chest X ray and use those images as a test on the already trained Res18 model but I got very bad results and accuray was dropped. Then I re train Res18 model with 500 new imagines keep the old images. Then testing was performed in which I was able to achieve accuracy of 98%.

Auther trained the Res18 model with 25 epochs, so changed the size of epochs to 30 again it aslo give good results.

I hope you like this blog. I’m sharing links for research paper, code and dataset.