Deep learning-driven feature engineering for lung disease classification through electrical impedance tomography imaging

Abstract

Thoracic imaging is vital for diagnosing lung diseases, as it provides a detailed visualization of the lungs. Despite significant advancements in medical imaging techniques, These methods pose critical challenges, such as high costs and the use of radiation in certain devices, which can raise serious concerns, limit accessibility, and increase potential health risks. Therefore main aim of this study addressing these issues by utilizing electrical impedance tomography (EIT), which is a non-invasive imaging technique that mitigates the risks of radiation exposure, reduces costs, and simplifies the interpretation of complex lung disease related patterns seen in traditional imaging methods. As EIT emerges as a promising imaging technique, this study investigates and develops a deep learning-based framework for classifying lung diseases using reconstructed EIT images. The proposed framework includes three feature extraction methods: Initial-Pretrained Weights Models (ResNet50 and DenseNet-201), fine-tuned convolutional 3D networks, and fine-tuned convolutional 3D accompanied by dense layer networks. Various machine models fed by extracted features were employed for lung sound disease classification both as individual learners and ensemble classifiers. The framework was evaluated on three classification tasks: binary classification (healthy vs. non-healthy) achieving 89.55% accuracy, 3-class classification (obstructive-related, restrictive-related, and healthy) achieving 55.29% accuracy, and 5-class classification (asthma, chronic obstructive pulmonary disease, interstitial lung disease, pulmonary infection, and healthy) achieving 44.54% accuracy. The proposed methods outperform state-of-the-art results and introduce novel approaches to EIT imaging classification.