🫁 Chest-X-Ray Classification

This project classifies chest X-ray images using a customized DenseNet121 neural network model.

Contents

• Technologies
• Dataset description
• Data preprocessing
• Proposed model
• Methodology
• Results
• Requirements
• Credits

Technologies

Dataset description

This dataset of chest X-ray (CXR) images was sourced from Kaggle.

It contains 3616 COVID, 6012 Lung Opacity, 10192 Normal, and 1345 Viral Pneumonia CXR images.

Every image was PNG, grayscale, and 299 x 299 pixels.

Data preprocessing

Images were processed using tf.keras.applications.resnet50.preprocess_input, which converts images from RGB to BGR and then zero-centers each color channel with respect to the ImageNet dataset without scaling. The images were also resized to 224 x 224 for ImageNet compatibility reasons.

In addition, real-time data augmentation was done. The table below details the parameters used.

For visualization purposes, some augmented pictures were saved to aug_images folder via the save_to_dir arg.

Proposed model

New bottleneck layers and a classifier head were added to a pretrained DenseNet121 model.

base_model = DenseNet121(
    include_top=False,
    weights='imagenet',
    input_tensor=Input(shape=(IMG_DIMS, IMG_DIMS, 3)),
    input_shape=(IMG_DIMS, IMG_DIMS, 3))

model = Sequential()
model.add(base_model)
model.add(Flatten())
model.add(BatchNormalization())
model.add(Dense(1024, activation='relu'))
model.add(Dropout(0.4))
model.add(BatchNormalization())
model.add(Dense(512, activation='relu'))
model.add(Dropout(0.4))
model.add(BatchNormalization())
model.add(Dense(256, activation='relu'))
model.add(Dense(4, activation='softmax'))

The base model layers were frozen, resulting in 52,141,828 trainable and 7,140,712 non-trainable parameters.

Methodology

The table below summarizes the dataset partition details.

Images and were trained for 25 epochs with a learning rate of 1e-4. The learning rate was lowered to 1e-5 for the last 5 epochs.

Results

Training Results

Testing Results

Output predictions were generated for the test input samples and compared to the true label values.

See sklearn.metrics.classification_report and sklearn.metrics.ConfusionMatrixDisplay for more details.

Requirements

Third party imports: matplotlib, numpy, IPython, PIL, sklearn, tensorflow, splitfolders

Credits

• https://www.kaggle.com/datasets/tawsifurrahman/covid19-radiography-database
• https://arxiv.org/pdf/1608.06993.pdf