Transfer Learning-Based Convolutional Neural Network for Six - Class Lung Disease Classification

Abstract

Convolutional neural network (CNN)-based deep learning (DL) is a promising solution in several applications, such as image classification and computer vision. Further, transfer learning-based CNN (TL-CNN) supports objective-specific applications, which can be achieved by integrating a pre-trained model such as Visual Geometry Group19 (VGG19) with a customized network. In the present work, convolution layers, ReLU functions, and pooling layers in the customized network are strategically chosen to classify six-class lung diseases: pneumonia, cardiomegaly, lung opacity, tuberculosis (TB), COVID-19, and normal. Also, the hyper parameters are optimized to refine the extended-VGG19 architecture (VGG19+customized network) in TL-CNN. The TL-CNN was built using Kaggle dataset of 3802 chest X-ray (CXR) images. Testing was conducted to evaluate the created TL-CNN, and the findings confirm that the proposed work outperformed existing models with 93.85% accuracy, 94.04% precision, 93.85% recall, 93.82% F1 score, and 99.56% area under the curve (AUC). Moreover, the confusion matrix of the TL-CNN is also discussed to access the distinguishability of the proposed multiclass classification.