Pneumonia Detection Using Deep Learning

1. Introduction

This project focuses on detecting pneumonia from chest X-ray images using deep learning techniques. The images are grayscale and belong to one of three classes:

• Class 0: No disease (healthy)
• Class 1: Bacterial pneumonia
• Class 2: Viral pneumonia

The dataset is derived from the publicly available Chest X-Ray Images (Pneumonia) dataset. The task involves building models to classify the images into the correct categories based on their features, and the performance is evaluated using accuracy, defined as:

Accuracy = (Number of correctly predicted test images) / (Total number of test images)

2. Data

The data is divided into two main folders:

• train_images contains 4,672 images, split as follows:
  • 1,227 images for class 0
  • 2,238 images for class 1
  • 1,207 images for class 2
• test_images contains 1,168 images for evaluation.

Additionally, a labels_train.csv file provides the corresponding labels for the training images.

3. Experiments

3.1 Architectures

Several CNN architectures, including ResNet-18, ResNet-34, and ResNet-50, were experimented with to build models capable of accurately classifying chest X-ray images. Due to the relatively small dataset size, smaller and less complex architectures, like ResNet-18 and ResNet-34, performed better than more complex models.

3.2 Data Augmentation

To improve model generalization and prevent overfitting, various data augmentation techniques were employed such as:

• Rotation
• Shear transformations
• Scaling
• Zooming
• Adjusting brightness and contrast

These augmentations helped diversify the dataset and improved the model's ability to generalize to unseen test data.

3.3 Learning Rate Scheduling

Two learning rate schedulers were explored:

• ReduceLROnPlateau: Decreases the learning rate when the validation accuracy plateaus, allowing finer adjustments during training.
• CyclicLR: Helps the model escape local minima by cyclically increasing and decreasing the learning rate.

3.4 Ensemble Models

Ensemble models combining different ResNet architectures were created to lower variance and leverage the strengths of different models. The ensemble models consistently outperformed individual models in terms of accuracy.

3.5 Automatic Image Generation

Automatic image generation was used to increase the number of images in each class to meet target class sizes, ensuring balanced datasets. For example, when the number of images per class was set to 50,000, additional images were generated using augmentations. Experiments were conducted with different target sizes such as 50,000, 30,000, and 20,000 images per class to assess the impact on model performance. This method of balancing the dataset through automated image generation helped enhance the model's ability to distinguish between classes.

4. Results

The results of the experiments indicate that the ensemble models performed better than the individual models in detecting pneumonia from chest X-ray images. Below is the table summarizing the performance of various models based on test accuracy:

Model Name	Architecture(s) Used	Accuracy (%)
V6EB3	ResNet-18 (V3R18), ResNet-50 (V3R50), ResNet-50 (V7R50)	86.472
V5EB2	ResNet-18 (V3R18), ResNet-50 (V3R50)	86.301
V7EB3	ResNet-18 (V3R18), ResNet-50 (V3R50), ResNet-50 (V1R50)	86.130
V2EB4	ResNet-18 (V3R18), ResNet-50 (V3R50), ResNet-50 (V7R50), ResNet-50 (V15R50)	85.787
V1R50	ResNet-50	85.445
V3R18	ResNet-18	84.589
V3R50	ResNet-50	84.246
V7R50	ResNet-50	83.732
V15R50	ResNet-50	83.732
V4R50	ResNet-50	83.39

Individual Model Details

Model Name	Samples per Class (*1000)	Batch Size	Learning Rate	Epochs Trained	Accuracy (%)
V1R50	30	9	0.0001	7	85.445
V3R18	30	32	0.001 CLR	14	84.589
V3R50	50	60	0.0001 RLRP	53	84.246
V7R50	30	120	0.0001	10	83.732
V15R50	30	32	0.0001 CLR	24	83.732
V4R50	50	9	0.0001 RLRP	9	83.39

The highest accuracy of 86.472% was achieved by the V6EB3 ensemble model, which combined ResNet-18 and ResNet-50 models. The V5EB2 ensemble model came close with an accuracy of 86.301%. Individual models like V1R50 (ResNet-50) achieved 85.445%, while V3R18 (ResNet-18) achieved 84.589%, demonstrating that smaller models can perform competitively with proper hyperparameter tuning.

5. Conclusions

The experiments showed that ensemble models significantly outperformed individual models in pneumonia detection from chest X-rays. Specifically, models combining ResNet-18 and ResNet-50 architectures yielded the highest accuracies. ResNet-18 demonstrated strong performance, especially with learning rate schedulers like ReduceLROnPlateau.

Data augmentation and learning rate scheduling played a critical role in improving model performance. In future work, it is planned to further enhance the models by exploring new hyperparameters, expanding the dataset, and potentially using transfer learning techniques. This project demonstrates the potential of deep learning in automating medical diagnostics for pneumonia detection.