MSDS19078_COVID19_DLSpring2020

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 used for this assignment can be found here.

https://drive.google.com/open?id=1-HQQciKYfwAO3oH7ci6zhg45DduvkpnK

Task 1 - A: Fine Tune only FC Layers of VGG16

In this task FC layers are replaced with 2 custom layers and all feature layers are freezed then trained model on given dataset.

Train Accuracy: 0.9414 | F1 Score: 0.9508
Validation Accuracy: 0.8980 | F1 Score: 0.9133
Test Accuracy: 0.9553 | F1 Score: 0.9626

Predictions

---

Task 1 - B: Fine Tune only FC Layers of ResNet18

In this task FC layers are replaced with 2 custom layers and all feature layers are freezed then trained model on given dataset.

Train Accuracy: 0.9028 | F1 Score: 0.9204
Validation Accuracy: 0.8767 | F1 Score: 0.8976
Test Accuracy: 0.9360 | F1 Score: 0.9470

Predictions

---

Task 2 - A: Fine Tune entire VGG16

In this task FC layers are replaced with 2 custom layers and all feature layers are unfreezed then trained model on given dataset.

Train Accuracy: 0.9751 | F1 Score: 0.9789
Validation Accuracy: 0.9260 | F1 Score: 0.9373
Test Accuracy: 0.9793 | F1 Score: 0.9825

Predictions

---

Task 2 - B: Fine Tune entire ResNet18

In this task FC layers are replaced with 2 custom layers and all feature layers are unfreezed then trained model on given dataset.

Train Accuracy: 0.9810 | F1 Score: 0.9840
Validation Accuracy: 0.9240 | F1 Score: 0.9360
Test Accuracy: 0.9713 | F1 Score: 0.9759

Predictions

---

Trained Models

Weights of trained models can be found here

https://drive.google.com/drive/folders/1wUg1gCygCSDe23Gnr8DKXVBpzHwPYACF?usp=sharing

Part 2

Dataset

Dataset used for Part 2 can be found here.

https://drive.google.com/open?id=1eytbwaLQBv12psV8I-aMkIli9N3bf8nO

Task 1 - A: Cross Entropy Loss

In this task two models VGG16 and ResNet18 were fully trained on given dataset using Cross Entropy Loss Function.

VGG16

Train Accuracy: 97.03 | F1 Score: 97.17
Validation Accuracy: 96.13 | F1 Score: 96.34

Training Confusion Matrix

Seperate confusion matrices are display for each class.

1. For COVID-19 class 182 samples were predicted correct.
2. For Normal class, 3913 samples were predicted correct.
3. For Pneumonia class, 2120 samples were predicted corrrect.

### Validation Confusion Matrix Seperate confusion matrices are display for each class. 1. For COVID-19 class 22 samples were predicted correct. 2. For Normal class, 392 samples were predicted correct. 3. For Pneumonia class, 213 samples were predicted corrrect.

Accuracy and Loss Curves

---

ResNet18

Train Accuracy: 97.48 | F1 Score: 97.53
Validation Accuracy: 95.28 | F1 Score: 95.46

Training Confusion Matrix

Seperate confusion matrices are display for each class.

1. For COVID-19 class 195 samples were predicted correct.
2. For Normal class, 3948 samples were predicted correct.
3. For Pneumonia class, 2101 samples were predicted corrrect.

### Validation Confusion Matrix Seperate confusion matrices are display for each class. 1. For COVID-19 class 22 samples were predicted correct. 2. For Normal class, 392 samples were predicted correct. 3. For Pneumonia class, 208 samples were predicted corrrect.

Accuracy and Loss Curves

---

Task 1 - B: Focal Loss

In this task two models VGG16 and ResNet18 were fully trained on given dataset using Focal Loss.

VGG16

Train Accuracy: 95.83 | F1 Score: 95.96
Validation Accuracy: 95.30 | F1 Score: 95.58

Training Confusion Matrix

Seperate confusion matrices are display for each class.

1. For COVID-19 class 176 samples were predicted correct.
2. For Normal class, 3879 samples were predicted correct.
3. For Pneumonia class, 2079 samples were predicted corrrect.

### Validation Confusion Matrix Seperate confusion matrices are display for each class. 1. For COVID-19 class 18 samples were predicted correct. 2. For Normal class, 392 samples were predicted correct. 3. For Pneumonia class, 210 samples were predicted corrrect.

Accuracy and Loss Curves

---

ResNet18

Train Accuracy: 96.11 | F1 Score: 96.23
Validation Accuracy: 94.93 | F1 Score: 95.10

Training Confusion Matrix

Seperate confusion matrices are display for each class.

1. For COVID-19 class 191 samples were predicted correct.
2. For Normal class, 3940 samples were predicted correct.
3. For Pneumonia class, 2032 samples were predicted corrrect.

### Validation Confusion Matrix Seperate confusion matrices are display for each class. 1. For COVID-19 class 22 samples were predicted correct. 2. For Normal class, 390 samples were predicted correct. 3. For Pneumonia class, 207 samples were predicted corrrect.

Accuracy and Loss Curves

---

Trained Models

Weights of trained models for Part 2 can be found here

https://drive.google.com/drive/folders/1wUg1gCygCSDe23Gnr8DKXVBpzHwPYACF?usp=sharing