Instruction for Preparing Fuzzy System Symposium Manuscript

Abstract

In constructing convolutional neural-network (CNN) models, it is important to select an optimization algorithm and learning rate according to the task. However, there have been no previous studies comparing the correctness rates of CNNs for pneumoconiosis image recognition. Therefore, in this study, we compare the accuracy of CNN models for classifying chest X-ray images using eight optimization algorithms and five different learning rates, and visualize the basis of the image judgment using Grad-CAM to clarify how the optimization algorithms affect learning and feature acquisition. The comparison was conducted using the highest accuracy for each optimization algorithm. The accuracy for the six optimization algorithms was approximately 90% on the validation dataset, while it was approximately 73% for Nadam and 49% for Ftrl, indicating that the accuracy varies depending on the optimization algorithm. The models with the highest accuracy, Adamax and RMSprop, focused mainly on the boundary of the lung field region. However, some of the other models that were close to the accuracy focused on the center of the lung field area, and it is necessary to examine the validity of their rationale. On the other hand, Nadam, which had a low accuracy, focused on only a part of the image and did not acquire enough features. Frtl focuses on the entire image regardless of lung field regions, and does not acquire features correctly. These results indicate that the selection of the optimization algorithm and the learning rate affect for the CNN-based detection models for the image of pneumoconiosis.