Atrous Convolutional Self-Attention-Based Capsule Network for Lung Disease Classification

Lung diseases remain a major global health concern, affecting the supply of oxygen to other parts of the body. There are several types of lung diseases, including asthma, COPD, and pneumonia. Numerous methods have been developed to identify these lung diseases; however, they still have several shortcomings, including long processing times, complex structures, and poor classification accuracy. To address these issues, a lung disease classification system is developed using the proposed method. First, pre-processing techniques are applied to improve image quality by reducing noise and enhancing contrast using Modified Histogram Equalization and Cross-guided Bilateral Filtering. The images are collected from the NIH ChestX-ray dataset. Next, Extended Lyrebird Optimization is utilized to select the optimal features, and the Squeeze-Excited DenseNet201 (SE-DenseNet201) model is employed for feature extraction. Finally, an Atrous Convolutional Self-Attention-based Capsule Network model is utilized for classification, and the Kookaburra Optimization Algorithm is employed for hyperparameter tuning. The proposed approach is evaluated using the NIH ChestX-ray dataset and achieves an accuracy of 92.70%, with 92.13% precision and 92.99% recall.