Performance Comparison of VGG16, VGG19 and Alexnet Pre-Trained Transfer Learning Architecture Models in the Convolutional Neural Network Algorithm in Classification of Lung Disease

Abstract

This study aims to comprehend the performance of transfer learning architectures (VGG16, VGG19, and Alexnet) in a Convolutional Neural Network for classifying lung diseases. Another objective is to determine the most superior transfer learning approach in this classification scenario. The dataset consists of 5 classes: normal lungs, pneumonia, bronchopneumonia, tuberculosis, and bronchitis. The data was sourced from Sinar Husni Deli Serdang Hospital through the radiology laboratory. The dataset was divided 80:20 for training and testing, with hyperparameters including a batch size of 32, 50 epochs, and optimization using Adaptive Momentum Optimization with a learning rate of 0.001. The research findings reveal that the VGG19 transfer learning architecture achieves the best performance with an accuracy of 59.17%, precision of 62%, recall of 59.2%, and an f-1 score of 58.8%. VGG16 ranks second with an accuracy of 55.83%, precision of 58%, recall of 55.8%, and an f-1 score of 55.2%. Alexnet has an accuracy of 49.17%, precision of 53.2%, recall of 49.2%, and an f-1 score of 50.6%. In an external test with 50 data points, VGG16 attains an accuracy of 54%, VGG19 scores 42%, and Alexnet records 46%. These models perform better in classifying normal lungs and tuberculosis compared to pneumonia, bronchopneumonia, and bronchitis. Analysis of lung image data demonstrates that homogeneity of RGB pixel values within a class supports transfer learning performance in classification. Conversely, heterogeneity in RGB pixel values can diminish the evaluation of that class.