Implementation of Convolutional Neural Network (CNN) MobileNetV2 in Lung Disease Classification from X-Ray Images

Abstract

The classification of lung diseases from X-ray images is often challenged by significant data imbalance, where minority classes like COVID-19 constitute only approximately 20% of the dataset compared to the majority classes. This condition can degrade model performance and introduce bias. This study aims to analyze the impact of data balancing strategies and training parameter variations to improve the accuracy of a Convolutional Neural Network (CNN) model based on the MobileNetV2 architecture. The experimental process systematically compared two learning rates (1e-3 and 1e-4) and two optimizers (Adam and RMSprop) across four distinct data handling scenarios: no augmentation, geometric augmentation only, the Mixup technique only, and a combination of both. The model was evaluated on a four-class X-ray image dataset comprising COVID-19, Normal, Pneumonia, and Tuberculosis. The optimal results were achieved by applying the combined approach of geometric augmentation and Mixup with a 1e-3 learning rate and the Adam optimizer. This configuration significantly outperformed other scenarios, reaching a testing accuracy of 96.62% and an average F1-Score of 96.63%, demonstrating excellent model generalization. This high-performing model has been successfully implemented in a mobile application using Flutter and TensorFlow Lite, serving as a practical tool to support the early diagnosis of lung diseases