Pneumonia Detection Project

Project Purpose

Detecting Pneumonia from X-Ray images using Convolutional Neural Networks

Dataset

The dataset contains 5,856 Chest X-Ray images. The images are split into a training set (train directory) and a testing set (test directory) of independent patients. There are two classes of images with labels NORMAL and PNEUMONIA. The dataset is publicly available at Labeled Chest X-Ray Images, which is itself taken from https://data.mendeley.com/datasets/rscbjbr9sj/3. under Licence: CC BY 4.0, and the main paper is https://www.cell.com/cell/fulltext/S0092-8674(18)30154-5

Requirements

• numpy
• seaborn
• pyplot from matplotlib
• tensorflow
• keras
  • Optimizer: Adam
  • Models: InceptionV3, ResNet50, DenseNet
• sklearn
  • Metrics: confusion_matrix, accuracy_score

Utils Directory

The utils directory includes Python files, each of which contains a CNN model and possibly some variants of it. It also contains helper.py, which includes the following auxiliary functions:

• datagen_train_dir
• datagen_test_dir
• plot_history
• plot_confusion_matrix
• plot_roc_au

Models Considered and Their Performance

1. Our Model:

   • Has only seven layers, including two convolutional layers and two maxpooling layers with only ~300,000 parameters
   • Shows the best accuracy and loss, the same as ResNet with 130 million parameters and other large pre-trained models
   • 95% accuracy

2. LeNet:

   • Quite old but produces satisfiable results
   • 91% accuracy

3. AlexNet:

   • Also quite old but produces satisfiable results
   • 91% accuracy

4. InceptionV3:

   • Better results than the previous models
   • 94% accuracy

5. ResNet with the following variants:

   • 5.1 With One Fully Connected Layer:
     • Best possible result among all models
     • 95% accuracy
   • 5.2 Nontrainable ResNet50 with One Fully Connected Layer:
     • Accuracy and loss values drop drastically
     • 83% accuracy
   • 5.3 Nontrainable ResNet50 with Several Normalisation Layers:
     • Improves accuracy and loss but still worse than the trainable base model (4.1)
     • 92% accuracy
   • 5.4 Nontrainable ResNet50 with Several Normalisation Layers and Further Data Preprocessing (e.g., rotation):
     • Makes the gap between data accuracy/loss and validation accuracy/loss much less
     • 86% accuracy

6. DenseNet:

   • 95% accuracy