Transfer Learning Based Multi-Class Lung Disease Prediction Using Textural Features Derived From Fusion Data

Abstract

Lung diseases pose significant challenges to public health worldwide, requiring accurate and efficient diagnostic methods for timely intervention. This research introduces a novel approach that uses deep learning algorithms on chest X-ray images generated from the fusion of individual datasets pertaining to Tuberculosis (TB), COVID-19, and Pneumonia to classify various lung illnesses. Employing the cutting-edge deep learning architectures such as MobileNetV2, Visual Geometry Group 16 (VGG16), InceptionNet, ResNet50, and EfficientNet, the usefulness of textural features derived using Local Binary Patterns (LBP) for improving classification performance is investigated. The EfficientNet and ResNet design achieves remarkable accuracy of 96.3% and 97.1% respectively on the fusion dataset. It is tested rigorously utilizing compound scaling to scale up the network in depth, width, and resolution simultaneously. This has resulted in better feature extraction and image representation. Additionally, methods for enhancing model performance are covered, such as feature engineering, data augmentation, and hyperparameter tuning. The findings demonstrate the potential of textural features-based deep learning techniques to effectively diagnose multiple lung diseases with subtle difference in original features from chest X-ray images. This research has a positive synergy in the clinical decision-making process and health care outcomes.

Keywords

• Chest X-ray images
• COVID-19
• EfficientNet
• InceptionNet
• local binary patterns (LBP)
• MobileNetV2