Abstract

This project was developed to participate in the "Student and Technical Progress" competition, aimed at fostering technological advancements. The inspiration for this idea stemmed from the pressing issue of improving medical education in my country, the Republic of Tajikistan. Considering the economic constraints of my nation, the affordability and accessibility of such software solutions are vital. This work represents a step towards leveraging technology to address healthcare challenges, particularly in medical diagnostics.

The project focuses on building a Computer Vision model for pneumonia detection using chest X-ray images. Harnessing the power of Convolutional Neural Networks (CNNs), the model classifies X-ray scans to identify signs of viral or bacterial pneumonia. The model was trained on a robust dataset sourced from Xian Djo University and Kaggle.com, undergoing meticulous preprocessing stages including data normalization and augmentation.

To present the results in an accessible format, a Streamlit-based web application was developed. This interface allows healthcare professionals to upload X-ray images and receive diagnostic insights, making the tool user-friendly and practical for real-world applications.

Technologies such as TensorFlow, Keras, and OpenCV were integral to the project, supporting neural network construction, image preprocessing, and visualization. The model's performance was evaluated using metrics like classification accuracy and confusion matrix analysis, ensuring its reliability in clinical settings.

This initiative underscores the potential of artificial intelligence in revolutionizing healthcare, offering a cost-effective, scalable, and innovative solution to improve diagnostic accuracy and medical education in resource-constrained regions.

Methodology

Data Preprocessing

1. Dataset Preparation:

• The dataset consisted of grayscale images resized to a uniform size of 150x150 pixels.
• Each image was normalized to have pixel values between 0 and 1 to ensure consistent input for the model.

def preprocess_image(image):
    image = cv2.resize(image, (150, 150))  # Resizing
    return image / 255.0  # Normalizing

2. Augmentation:

Image augmentation techniques such as horizontal flipping, rotation, and zoom were applied using the ImageDataGenerator from Keras to increase the diversity of the training set and reduce overfitting.

from keras.preprocessing.image import ImageDataGenerator

datagen = ImageDataGenerator(
    rotation_range=20,
    width_shift_range=0.2,
    height_shift_range=0.2,
    horizontal_flip=True,
    zoom_range=0.2
)

3. Splitting:

The dataset was split into training, validation, and test sets with proportions of 70%, 20%, and 10%, respectively.

Model Architecture

Base Model

A Convolutional Neural Network (CNN) was implemented using the Keras Sequential API.

Layers Configuration

1. Five convolutional layers with ReLU activation for non-linear feature extraction.
2. Batch normalization layers to stabilize and accelerate training.
3. MaxPooling layers to downsample feature maps after each convolutional block.
4. Dropout layers to mitigate overfitting.

from keras.models import Sequential
from keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout, BatchNormalization

model = Sequential([
    Conv2D(32, (3, 3), activation='relu', input_shape=(150, 150, 1)),
    BatchNormalization(),
    MaxPooling2D(pool_size=(2, 2)),
    Dropout(0.25),
    Conv2D(64, (3, 3), activation='relu'),
    MaxPooling2D(pool_size=(2, 2)),
    Flatten(),
    Dense(128, activation='relu'),
    Dropout(0.5),
    Dense(1, activation='sigmoid')
])

---

Model Compilation

Loss Function

Binary Crossentropy was used as the loss function, suitable for binary classification tasks.

Optimizer

RMSprop was chosen for optimization due to its effectiveness in training deep networks.

model.compile(
    optimizer='rmsprop',
    loss='binary_crossentropy',
    metrics=['accuracy']
)

Evaluation Metric

Accuracy was the primary metric used to evaluate the model's performance.

---

Training

Batch Size and Epochs

• Batch size: 32
• Number of epochs: 12

Learning Rate Scheduler

The ReduceLROnPlateau callback was implemented to reduce the learning rate when validation accuracy plateaued.

from keras.callbacks import ReduceLROnPlateau

lr_scheduler = ReduceLROnPlateau(
    monitor='val_accuracy',
    factor=0.3,
    patience=2,
    min_lr=1e-12
)

Validation

Validation was performed after each epoch using the validation dataset to monitor generalization performance.

---

Evaluation

Test Performance

The trained model was evaluated on the test set to compute the loss and accuracy.

Classification Report

Metrics such as precision, recall, and F1-score were calculated for each class to gain deeper insights into the model’s performance.

# Example: Generating a classification report
from sklearn.metrics import classification_report

predictions = (model.predict(test_images) > 0.5).astype("int32")
print(classification_report(test_labels, predictions, target_names=["Pneumonia", "Normal"]))

Visualizing Results

• Learning curves were plotted to analyze training and validation performance over epochs.
  

• Confusion matrix plotted for a detailed class-wise performance summary.

from sklearn.metrics import confusion_matrix
import seaborn as sns
import matplotlib.pyplot as plt

cm = confusion_matrix(test_labels, predictions)
sns.heatmap(cm, annot=True, fmt='d', cmap='Blues', xticklabels=["Pneumonia", "Normal"], yticklabels=["Pneumonia", "Normal"])
plt.xlabel("Predicted")
plt.ylabel("Actual")
plt.show()

Results

1. Model Performance on Validation Data

Metric	Value
Training Accuracy	96.66%
Validation Accuracy	75.00%
Validation Loss	0.4328

---

2. Test Dataset Evaluation

Metric	Value
Test Accuracy	90.22%
Test Loss	0.2470

---

3. Classification Report

Class	Precision	Recall	F1-score
Pneumonia (Class 0)	93%	92%	92%
Normal (Class 1)	86%	88%	87%

Overall Metrics

Metric	Accuracy	Macro Average	Weighted Average
Overall Metrics	90%	90%	90%

---

4. Confusion Matrix

	Predicted Pneumonia	Predicted Normal
Actual Pneumonia	358	32
Actual Normal	28	206

Useful Links

• Streamlit Application (Ready for Usage):
  Pneumonia Detection App