Deep learning model to classify Chest X Ray scan images between 3 classes including Viral pneumonia and COVID-19.

Project Statement This Coursera project aims to classify XRay scan images between 3 classes: Normal, Viral pneumonia and COVID-19. The project utilizes a dataset of images which can be found at Kaggle here. The dataset contains a total of 2,924 grayscale images with a significant class imbalance due to under-representation of the COVID-19 class.

• 1,341 normal images
• 1,345 viral pneumonia images
• 219 COVID-19 images

Working

• I first split the dataset so that to reserve a validation set of 30 images per classes (90 images total). This will allow to measure the accuracy of the model during training.

• I use ResNet18 network, with weights pre-trained on ImageNet, only adjusting the final classification layer to predict between the 3 classes.

• I allow to fine-tune all model parameters during training

• Training conditions:

  • Adam optimizer with learning rate of 3e-5
  • Pytorch cross entropy loss function
  • Batch size of 6
  • Resized images to (224,224) and normalized as per model requierements

• The model converges in less than one epoch and achieves an accuracy of over 98% on the validation images

98% accuracy on average with 100% accuracy on COVID detection.