Leveraging DenseNet Features for Machine Learning Based Lung Disease Diagnosis From X-Rays

Infectious lung diseases, including COVID-19, bac- terial pneumonia, and viral pneumonia, are significant concerns for public health worldwide. Chest X-rays (CXR) are a com- mon diagnostic tool, but their manual interpretation can be complicated due to the similar visual characteristics of these diseases. This paper presents an automated system that utilizes features from deep learning models (DenseNet-121, DenseNet- 169, DenseNet-201) to classify lung diseases into four distinct categories: bacterial pneumonia, COVID-19, viral pneumonia, and normal. With a dataset of 4,202 X-ray images, DenseNet-201 achieved the highest accuracy of 94.10%, showing its potential to assist radiologists in clinical diagnosis. Additionally, integrating features from DenseNet models with machine learning classifiers such as SVM and Random Forest improved overall test accuracy, with DenseNet-201 combined with SVM achieving a near-perfect test accuracy of 97.67%. Feature extraction from the penultimate layers of DenseNet models played a critical role in enhancing the classification process, allowing traditional machine learning classifiers to achieve higher accuracy and reliability across various lung disease categories.