Advancing Diagnostic Accuracy in Lung Disease Severity Classification Using Multi-Domain Features

Abstract

Accurate classification of lung diseases is crucial for early diagnosis and effective treatment. This study presents an optimised classification framework that utilises multi-domain feature extraction and a deep neural network (DNN) for categorising lung disease severity from CT scan images. The dataset collected from a local hospital includes 266 CT scans of Lung cancer, COVID-19, and pneumonia, categorized into mild (43 images), moderate (82 photos), and severe (141 images) cases. To address class imbalance, the synthetic minority oversampling technique (SMOTE) was applied, ensuring equal representation across categories. A total of 30 multi-domain features were extracted using a comprehensive feature extraction methodology that combined wavelet packet decomposition (WPD) with statistical, texture, shape, edge detection and Grey Level Cooccurrence Matrix (GLCM) features. These features captured diverse spatial and frequency-based characteristics of lung disease patterns, enabling robust model input. This study focuses on classification based on the severity of patient condition within three different classes: Mild, Moderate, and Severe, related to lung disease. The classification was performed using a Deep Neural Network (DNN) with fine-tuned hyperparameters. The model achieved a training accuracy of 95%. The findings underline the potential of this approach in improving automated diagnostic systems. The extracted features provide a comprehensive representation of disease patterns, while the DNN leverages these features for precise classification. This methodology offers valuable insights for applications in medical imaging. This research contributes to the field of medical image analysis by integrating robust feature extraction techniques with advanced classification models, paving the way for more accurate and reliable lung disease diagnosis.