Chest-Classification program

Author: Shiwei Liu, College of Artificial Intelligence, Nankai University

E-mail: 2011832@mail.nankai.edu.cn

About this program

This project is my final assignment for the course "Artificial Intelligence Technology" during my junior year at Nankai University.It is used to classify whether a given human lung X-ray image is diseased or not. I built a CNN by Pytorch to deal with this problem and it works well in this dataset (please download it in./chest_xray). After only 100 epochs training, it achieves 88% accuracy in validation dataset and 75% accuracy in test dataset. I also plot the "train and val loss" curve and test-results images.

Requirements

1. Create a conda environment by running this command(replace YOUR_ENV_NAME):

   conda create -n YOUR_ENV_NAME python==3.9 
   

2. Install Pytorch which is suitable for your GPU or CPU(not recommended):

   pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
   

3. Use pip tool to install other required package:

   pip install numpy
   pip install os
   pip install matplotlib
   

Run this program

1. Clone this project to your own PC:

   git clone 
   

2. Download dataset, and place file chest_xray in ./Chest_Classify
3. Activate the conda environment created before conda activate YOUR_ENV_NAME and run this command (replace NUM_OF_EP):

   python main.py --epochs=NUM_OF_EP
   

   Trained module will be saved as ./model.pt.

Result

When main.py finished the train function I use matplotlib to plot the curve of "train loss/accuracy" and "validation loss/accuracy":

You can write your own test.py to visualize the training result.

Further discussion

The overfitting situation is very serious. Maybe more dropout or a simpler network structure?