Detection of Pneumonia Using Deep Learning Approach from X-Ray Images
No Thumbnail Available
Date
2024-09-27
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Pneumonia classification is developing an automated system capable of distinguishing between three distinct categories—normal, non-pneumonia lung diseases, and pneumonia—using chest X-ray images. Pneumonia is a severe lung infection that can lead to significant morbidity and mortality if not promptly diagnosed and treated. However, the challenge arises from the fact that other lung conditions, like chronic obstructive pulmonary disease (COPD) or tuberculosis, can present with symptoms and X-ray findings similar to pneumonia, making accurate differentiation crucial. The objective of this study is to design and evaluate a machine learning model that can accurately classify these categories, thereby aiding in the early and accurate diagnosis of pneumonia versus other conditions. This classification task is vital for improving clinical outcomes, reducing diagnostic errors, and ensuring that patients receive appropriate treatment as quickly as possible.
The primary objective of this research is to develop an effective model for pneumonia classification using deep learning approach. By categorizing chest X-ray images into three classes—non-pneumonia (other lung diseases), normal (healthy lungs with no signs of disease), and pneumonia—we aim to enhance diagnostic accuracy and improve radiologist performance. To achieve this, we utilize a data set collected from Tikur Anbessa Specialized Hospital, comprising 3000 chest X-ray images, with 1000 images per class to ensure balanced representation before augmentation; and also, we used a ratio of 80:10:10 (80% training, 10% validation, and 10% testing) splitting ratio.
We employed an experimental research approach, selecting three state-of-the-art pretrained models—InceptionV3, ResNet50, and VGG16—for transfer learning, alongside constructing a custom CNN-L5 model. Through a series of experiments, we investigated various image prepossessing techniques, including re-sizing, normalization, and image enhancement, to optimize model performance. We explored different combinations of epochs, batch size, and learning rates for all models, while also experimenting with fine-tuning the pretrained models. The most effective model for pneumonia classification was then selected based on these trials. The experimental results showed that the CNN-L5 model, trained on enhanced image data sets with 30 epochs, a batch size of 64, and the inclusion of dropout, achieved superior performance with a classification
accuracy of 96.8%. This highlights the importance of using appropriate preprocessing techniques and optimizing model architecture to achieve high performance in pneumonia classification.
However, this study was limited to data from a single hospital, which may not represent the diversity of patient populations, imaging techniques, and conditions found in other healthcare settings. Consequently, the model's performance may not generalize well across different environments. To address this limitation, future work should incorporate data from multiple hospitals to improve the model's robustness and broader applicability.
Description
Keywords
Pneumonia Classification, Deep Learning, Convolutional Neural Network, Digital Image Processing