Hybrid Deep Learning Approaches for Early Diagnosis of Chronic Lung Diseases

The chronic obstructive pulmonary disease (COPD) impacts the health of people worldwide as it either increases mortality or morbidity rates or both among patients suffering from this disease. Early diagnosis is an important factor that would help to increase patient survival rates and the quality of service. We propose a novel system that utilizes a multimedia dataset of interstitial lung diseases (ILDs) from the University Hospitals of Geneva (HUG), publicly available for research purposes. In the training process, 80 percent of the data was used to build a classifier; 20 percent of the data was used in the testing process. To improve the dataset size, we used data augmentation techniques including rotation, flipping, and scaling. Feature extraction was then performed using both manual methods like HOG and GLCM, as well as fully automated methods such as DenseNet 121 and EfficientNet B6 on the augmented data. The results of these are concatenated in different combinations and fed to classifiers such as SVM, XGBoost, LightGBM, KNN and LR. Comparisons are made with respect to accuracy, precision, recall and $\mathbf{f 1}$ score. Finally, results derived from combination formed by concatenating features extracted from DenseNet 121 and HOG are proving efficient performance and are giving an accuracy of 98.66 percent with classification on Support Vector Machine.