Covid-19 classification using Chest X-Ray Dataset

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

Part 1

Dataset

Dataset is available here. This dataset contains chest X-Ray images classified into infected and normal categories.

Description

Classification was done with transfer learning using the VGG-16 and ResNet-18 architectures and PyTorch. Experiments were done using two approaches:

1. Using pre-trained feature extraction layers: Convolution layers pre-trained on ImageNet were used to extract features and custom dense layers were added at the end of the network and trained.
2. Fine-tuning: Models trained on ImageNet were finetuned. Three different configurations were used in this approach.
   1. In the first configuration, only one convolution layer including the dense layers were trained.
   2. A few more convolution layers were trained.
   3. The entire network was finetuned. Note that the results section contains the results of this configuration only.

Experimental Setup

Both the models were modified to include only two final dense layers containing 120 and 2 units respectively.

Additionally, in all the experiments the models were trained for six epochs with Adam optimizer (learning rate: 0.001, beta1: 0.9, beta2: 0.999, epsilon: 1e-8)

Training Statistics

Loss and accuracy curves during training for all the experiments are presented below

VGG-16

Layers Trained	Train Accuracy	Train Loss	Validation Acuracy	Validation Loss
Dense layers	0.916	0.216	0.891	0.302
One Convolution layer + Dense layers	0.954	0.128	0.896	0.292
Three Convolution layers + Dense layers	0.943	0.165	0.910	0.223
Entire Network	0.930	0.190	0.896	0.239

ResNet

Layers Trained	Train Accuracy	Train Loss	Validation Acuracy	Validation Loss
Dense layers	0.881	0.291	0.859	0.339
One Convolution block + Dense layers	0.957	0.114	0.895	0.318
Two Convolution blocks + Dense layers	0.957	0.113	0.912	0.255
Entire Network	0.944	0.154	0.908	0.244

Results

VGG-16

Pre-trained Feature extraction

Accuracy	F1-score
0.953	0.961

Confusion Matrices

• Train

  

• Validation

  

• Test

  

Fine-tuning

Accuracy	F1-score
0.973	0.978

Confusion Matrices

• Train

  

• Validation

  

• Test

  

ResNet-18

Pre-trained Feature extraction

Accuracy	F1-score
0.925	0.938

Confusion Matrices

• Train

  

• Validation

  

• Test

  

Fine-tuning

Accuracy	F1-score
0.968	0.973

Confusion Matrices

• Train

  

• Validation

  

• Test

  

Analysis

Higher accuracies were acheived when the entire network was finetuned. This can be due to the fact that the original ImageNet dataset is different from the current task involving medical images. Fine-tuning of the convolution layers adjusted the features extracted from the network and showed better results.

Part 2

Multilabel classification of x-ray images into {normal, pneumonia, covid} labels. The dataset is unbalanced for the covid class

Dataset

Dataset is available here

Results

VGG-16 (Cross Entropy Loss)

Subset Accuracy	Macro-averaged F1-score
0.952	0.899

Train Set

Validation Set

VGG-16 (Focal Loss)

Subset Accuracy	Macro-averaged F1-score
0.952	0.920

Train Set

Validation Set

ResNet-18 (Cross Entropy Loss)

Subset Accuracy	Macro-averaged F1-score
0.957	0.913

Train Set

Validation Set

ResNet-18 (Focal Loss)

Subset Accuracy	Macro-averaged F1-score
0.935	0.882

Train Set

Validation Set

Weights

Weights can be found here