COVID-19 classification from chest X-rays using Deep learning models

This repository contains code and results for COVID-19 classification project by Deep Learning Spring 2020 course offered at Information Technology University, Lahore, Pakistan. This project is only for learning purposes and is not intended to be used for clinical purposes.

Dataset

Access Dataset here

Class	Train	Validation	Test
Infected	4919	615	615
Normal	7081	885	885

Models

Access VGG16 Model here

Access ResNet18 Model here

Covid-19 X-rays images classification using Transfer learning

Experiments Summary

Epochs = 10

Learning Rate = 0.0001

Model	Parameters	Training Time	Train	Validation	Test
VGG16	2 FC layers	<3 hours	91.33	91.33	97
ResNet18	2 FC layers	<1 hrs 30m	84	84	92
VGG16	Features.28 + 2 FC layers	<3 hours	88	89	95
ResNet18	Layer4 + 2 FC layers	<1 hrs 30m	61.72	61.77	61
VGG16	Features 24,26,28 + 2 FC layers	<3 hours	90	89	96
ResNet18	Layer3,4 + 2 FC layers	<2 hours	66.72	67.77	67
VGG16	All Features + 2 FC layers	<3 hours	91	91.77	96
ResNet18	All Blocks + 2 FC layers	<2 hours	65	67.77	67

VGG16 vs Resnet18

Model Evalualuation

Discussion :

VGG gives more accuracy than Resnet but it requires more training time than Resnet. It is costly in terms of parameters so it provides that much high accuracy easily. Resnet requires less training time than VGG but due to less parameters it is not able to provide that high accuracy in comparison to vgg. But setting optimal parameters like learning rate we can tune this to work better.

Dealing with high class imbalancement using Focal Loss

Dataset

Access Dataset here

Class	Train	Validation
Covid-19	200	28
Pneumonia	2000	200
Normal	4000	400

we have 629 unlabeled testing images

Models

Access VGG16 Model here

Access ResNet18 Model here

Experiments and Analysis