Comparison Analysis of Convolutional Neural Networks Using CPU and GPU in the Diagnosis of Lung Diseases (#1253)

Authors

Huamaní Navarrete, Pedro Freddy

Maldonado Lezama, Lisset Fernanda

Moreano Rojas, Eder Omar

Abstract

This article describes the comparative analysis of the use of CPU and GPU in the training and validation of three convolutional neural network models, to diagnose lung diseases from central chest X-rays. Since tuberculosis and Covid-19 present similar symptoms, it is possible to confuse the diagnosis and provide incorrect treatment. For this reason, two convolutional network models, InceptionV3 and VGG16, and an arbitrary one consisting of ten hidden layers, were chosen to compare them and thus achieve a more accurate diagnosis. Likewise, the JupyterLab interface was used with the Python programming language, complemented by the TensorFlow and Keras libraries. Then, for the training stage, 2,535 images were used, and the transfer learning technique was applied using the computer's CPU and GPU, with the purpose of analyzing the effectiveness when comparing each of the presented cases; likewise, this group of images also included the diagnosis of healthy patients. Regarding the evaluation, the metrics Accuracy, Recall, F1-Score, and General Precision were used to identify the performance of the arbitrary network model when compared to the other mentioned models. In this way, the arbitrarily proposed convolutional neural network model achieved higher accuracy, equivalent to 92.70% when the CPU was used, while when the GPU was used, this accuracy increased to 94.28%.