Classification method for nailfold capillary images using an optimized sugeno fuzzy ensemble of convolutional neural networks

Abstract

This study developed a novel binary classification method for analyzing nailfold capillary images associated with the risk of developing sclerosis. The proposed approach combined a Sugeno fuzzy integral inference system with an ensemble of convolutional neural networks (CNNs), including GoogLeNet, ResNet, and DenseNet. Nailfold capillary images are highly valuable for diagnosing and monitoring various systemic diseases. They can reveal early indicators of systemic sclerosis, such as capillary enlargement, loss, or hemorrhages. The study obtained nailfold capillary images from a hospital in Taiwan, with 80 % allocated for model training and the remaining 20 % reserved for testing purposes. The proposed method achieved a high performance with an accuracy of 85 %, a recall of 81.82 %, a precision of 90 %, and an F1 score of 85.17 %. In comparison, individual CNN models (GoogLeNet, ResNet, and DenseNet) achieved accuracies of 73.33 %, 67.96 %, and 70.83 %, respectively. These results demonstrate that the proposed integrated method outperforms single-model approaches in classifying nailfold capillary images more accurately and efficiency. Using CNN models as a novel application opens new avenues for research in related image analysis fields.