urvil38 / big-data-project-winter-2024 Goto Github PK

Our project focuses on creating a robust and reliable system designed to classify X-ray images of lungs into distinct classes, including COVID-19, viral pneumonia, lung opacity, and normal cases. The data for this project is organized into three key categories: training, validation, and test sets, ensuring comprehensive learning and evaluation phases. This endeavor aims to significantly enhance healthcare delivery and improve patient outcomes by leveraging a dataset partitioned into training, validation, and test sets. By distinguishing between these conditions effectively, our project seeks to enhance medical diagnostic processes significantly.

The project is structured around two core objectives:

1. Data Preparation: Compiling a comprehensive dataset of labeled X-ray images. This involves collecting, preprocessing, and augmenting images to ensure a diverse and representative dataset for training our models.

2. Model Training and Evaluation: Training and fine-tuning ResNet and EfficientNet models on our dataset. The goal is to compare these models in terms of accuracy, efficiency, and their ability to generalize from training data to unseen X-ray images.

• Variability in X-ray Images: Addressing differences in image quality, patient positioning, and visible artifacts to ensure consistent model performance.
• Class Imbalance: Managing uneven distribution of classes within the dataset, particularly the lower prevalence of COVID-19 cases, to avoid model bias.

Dataset: https://www.kaggle.com/datasets/tawsifurrahman/covid19-radiography-database

In Order: COVID-19, Viral Pneumonia, Lung Opacity, Normal

Each class has over 5000 X-ray images:

Our preprocessing workflow involves resizing images to a consistent dimension, normalizing pixel values, and applying data augmentation (e.g., rotation, flipping) to enhance model training.

• ResNet: We will utilize the ResNet architecture, known for its deep residual learning framework, to capture complex patterns in the data.

• EfficientNet: EfficientNet will be used for its scalable architecture, allowing us to efficiently manage model complexity and computational cost.

1. ResNet vs. EfficientNet Performance: How do ResNet and EfficientNet compare in terms of accuracy and computational efficiency for pneumonia detection in our dataset?

2. Preprocessing and Augmentation Impact: Which preprocessing and data augmentation techniques are most effective for each model in improving generalization to new images?

3. Class Imbalance Mitigation: What strategies are most effective in balancing the class distribution for both models, ensuring accurate detection across all pneumonia types?

By addressing these questions, our team aims to develop a reliable and efficient tool for pneumonia detection, offering valuable support to healthcare professionals.