A Deep Learning Model to Diagnose Lung Disease using Genetic Algorithm Based Feature Selection Method

Abstract

Chest X-ray (CXR) image view information can help to improve and strengthen a computer-aided-diagnosis (CAD) system. The goal of this work is to develop a deep-learning model with an effective feature selection method for early diagnosis of pneumonia lung infections using CXR images. Here, a Genetic Algorithm based feature selection method is embedded in an improved Residual Network model with 53 layers, which is a variation of the ResNet-50 model. The model is evaluated using the Kermany dataset, which has 5856 CXR images. 80% of the photos in this image collection are used for training, while 20% are utilized for testing. Python is used for implementation to illustrate the efficacy of the suggested paradigm. The performance results are analysed and compared to models that have already been trained, such as GoogLeNet, ResNet18, and DensNet121. Findings: On the Kermany dataset, the suggested model achieves 98.1% accuracy, 98.2% sesitivity, 97.9% specificity and 97.6% precision, which is superior to the most advanced models addressed in the literature. We have added a feature selection layer in improved ResNet50+3 layer architecture. These three extra layers solve the vanishing gradient problem in the ResNet50 architecture, making it easier to train and adding feature selection layer to improve accuracy. According to the findings of the whole inquiry, the proposed model not only outperforms most classifiers in terms of accuracy, precision, and recall, such as GoogLeNet, ResNet50, and DenseNet121, but it is also a very adaptable model that works well on a variety of datasets.