EE416 Final Project

---

A convolutional neural network that classifies a chest CT image of various patients as having COVID-19 pneumonia, other pneumonia, or being healthy.

Overview

Dataset

• ~4,400 2D lung CT images
• 3 classes/labels
  • COVID-19 pneumonia
  • other pneumonia
  • healthy condition

Test & Training Data

• split 80% training and 20% test using 5-fold cross validation
• (optional) data augmentation by rotating images 90, 180, and 270 degrees to quadruple the amount of images

Tasks

1. create a program that splits the dataset into training and test data and automatically forms a folder structure
   • 80% training, 20% test using 5-fold cross validation
2. finish the convolutional neural network (CNN) by implementing the 2nd layer convolution kernels, convlution operator, and summation
   • determine amount of parameters in this CNN
3. train the model using the training dataset
   • decide model architecture
   • decide training loss function
   • decide optimizer
   • decide training parameters
4. test the model on the test dataset
   • outputs an accuracy
   • target: >85% (partial credit: >75%)
5. create report, which includes:
   • detail of CNN model and hyperparameters used
   • graphs of training and test loss across epochs (iterations)
   • accuracy of best model on the test set

Notes:

FCN: fully-connected network CNN: convolutional neural network

Running

Before running the program, place the folder of images labeled "COVID" in Data/Covid, the folder of images labeled "Healthy" in Data/Healthy, and the folder of images labeled "Other pnumonia" in Data/Others. The program will automatically move and rename these files.

Then, you can run: python eg.py

Results

Training times for various numbers of epochs (CUDA w/ RTX 3070)

# Epochs	Training & Validation Time (seconds)
1	11.818
10	88.611
25	250.234
50	466.927

Accuracys for various numbers of epochs

1 epoch (average = 52.0%):

10 epochs (average = 76.5%):

25 epochs (average = 80.3%):

50 epochs (average = 89.5%):