Investigating And Developing Lung Diseases Classification Model Using Ensemble Deep Learning

Abstract

Lung diseases, caused by the COVID-19 pandemic, pose a significant risk to millions of people. Chest X-ray imaging is widely used for diagnosing lung diseases, but accurate diagnosis remains challenging due to shortages of trained radiologists. A computer-aided recognition system has been proposed to minimize errors, using an ensemble of CNNs. In this research, we present convolutional neural network-based ensembles for classifying chest X-ray images into five classes: Pneumonia, Pneumothorax, Tuberculosis (TB), COVID-19, and normal. To minimize misclassification, we combined three procedures: Balance class, Image augmentation techniques with Keras ImageDataGenerator class & using an ensemble model with transfer learning, three separate CNNs—VGG-16, ResNet-50, and MobileNetV2—are combined to create a picture categorization system. The system trained and tested using 7340 chest X-ray images data type from the National Institute of Health chest X-ray repository and Jimma University Medical Center radiology department, significantly reduces manual visual + errors and can serve as a decision support for physicians. We used 80/20 by splitting the data into 80% for training, 10% for tests, and 10% for validation to train each three models namely VGG-16, MobileNetV2, and ResNet-50 and then we trained the concatenate ensemble of the three models. We compared the results with each other and finally compared them with the concatenated ensemble of the three models. As we compared to the state-of-the-art methods the promising classification performance of our proposed method achieved an accuracy of 97.02% meaning that our model achieved 4.29% more accuracy than the benchmark. While the accuracy of MobileNetV2 is 92.05%, VGG16 is 95.73% and ResNet50 is 89.20% so the high accuracy we obtained is by ensemble which is 97.02%. An ensemble of CNNS models, despite higher computational and modeling costs, offers superior performance and robustness in lung disease classification, outperforming individual models and enhancing classification accuracy