Determining the highest accuracy among the Convolutional Neural Network architectures: VGG16, ResNet50, or MobileNet for Pneumonia detection in 2023: The Determination of the Best Convolutional Neural Network Architecture: VGG16, ResNet50, or MobileNet for Pneumonia Detection in 2023

• Autores: Enciso-Ortiz, Sergio Elías, Mamani-Vilca, Ecler, Ordoñez-Ramos, Erech
• Formato: artículo
• Fecha de Publicación: 2023
• Institución: Universidad Nacional Micaela Bastidas de Apurímac
• Repositorio: UNMB-Riqchary
• Lenguaje: español
• OAI Identifier: oai:revistas.unamba.edu.pe:article/32
• Enlace del recurso: https://revistas.unamba.edu.pe/index.php/riqchary/article/view/32
• Nivel de acceso: acceso abierto
• Materia: Arquitecturas de Redes Neuronales Convolucionales; MobileNet; ResNet50; VGG16; Convolutional Neural Network Architectures

Convolutional Neural Networks (CNN) are used for the recognition of X-ray images and other applications. Currently, there are studies comparing the effectiveness of CNN architectures such as VGG16, ResNet50, and MobileNet with different input parameters during training, creating uncertainty among developers of image classification applications. We applied identical inputs for the training of the CNNs under study to address this lack of information. To address this, a Kaggle database consisting of 5856 images was utilized. From this database, a systematic sample of 746 lung X-ray images, both healthy and with pneumonia, was selected. To ensure image normalization, tools like iloveimg and ReNamer were employed. Furthermore, Python was used with Google Colab and various libraries including tensorflow, matplotlib, numpy, os, cv2, and random to execute the different architectures. The methodological design was based on a quantitative approach, utilizing comparison tables and the images acquired from the Kaggle database. The results obtained indicated that the accuracy percentage was 89.83% for VGG16, 91.82% for ResNet50, and 80.21% for MobileNet, leading to the conclusion that ResNet50 is the most accurate architecture in this context.