Enhanced Diagnosis of Chest X-Ray Using Hybrid Deep Learning Models and Feature Selection Techniques

Abstract

Background: Chest X-ray (CXR) images are important for diagnosing lung diseases such as pneumonia and tuberculosis. They help medical professionals determine the functions of the heart and lungs. The lungs may change as a result of certain heart issues, and specific disorders may cause anatomical alterations of the heart or lungs. Objectives: To enhance the classification accuracy of CXR images, particularly for identifying eight types of abnormal lung lesions, by applying advanced feature selection and fusion techniques in combination with deep learning models. Materials and Methods: This study utilized a dataset of CXR images with normal and abnormal classes; however, we had only eight lesions in the abnormal class. Initially, we preprocessed the images to improve their quality and suitability for further analysis. We then modified two deep learning models, ResNet50 and DenseNet201. Transfer learning (TL) techniques were employed for training. The features extracted from these models were retained separately. We utilized a parameter-optimized ant colony optimization (ACO) algorithm to refine the feature selection, and the features from both models were fused into a single feature set. Meanwhile, at the preprocessing step, we used five additional statistical methods, including the Kruskal-Wallis test, ReliefF, analysis of variance (ANOVA), chi-square test, and minimum redundancy maximum relevance (MRMR), to identify the most important features, separately from the above process. The fusion is then enforced after obtaining important features from two different processes to enhance the efficiency of our architecture. We then utilized various machine learning classifiers. Results: The binary classification between normal and abnormal achieved 95.9% ± 0.5% accuracy, 95.95% ± 1.09% sensitivity, 93.91% ± 0.34% specificity, 93.91% ± 0.37% precision, and 94.85% ± 0.56% F1 score on the hybrid approach. Multiple classifications of the eight abnormality lesions revealed a promising average area under the curve (AUC) value of 0.872. Conclusion: The combination of deep learning models with advanced feature selection methods is beneficial for not only improving the classification outcomes but also ensuring accuracy and validity.