The anomaly detection mechanism using deep learning in a limited amount of data for fog networking

• Published in: Computer communications Vol. 170; pp. 130 - 143
• Main Authors: Horng, Gwo-Jiun, Liu, Min-Xiang, Hsu, Chien-Chin
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 15.03.2021
• Subjects: AutoEncoder; Deep learning; Generative Adversarial Network; Deep learning; AutoEncoder; Generative Adversarial Network
• ISSN: 0140-3664; 1873-703X
• DOI: 10.1016/j.comcom.2021.01.036

The treatment of brain ischemia is the use of tissue plasminogen activator. But after treatment there may be a risk of Brain hemorrhage. Under the pressure of rescue time, medical personnel and their families must make decisions. There is currently no benchmark that can provide what can happen after treatment. This is because there are too many uncertainties. This study proposes that an adaptive deep autoencoder model is used to learn the features of a particular data and analyze the results that the data belongs to. For data preprocessing, the study also proposes methods such as K-means and connecting and labeling non-background areas to de-noise the image and preserve the desired areas to the maximum extent. In addition, we use Variational AutoEncoder Wasserstein Generative Adversarial Network with Gradient Penalty (VAE WGAN-GP) to generate 3D medical images to solve the problem of too few training data. We study key techniques such as pre-processing of real data and image generation with limited real data, so that the model can be trained smoothly. From our experiments, we found that the autoencoder model analyzes real data that is bleeding and non-bleeding and achieves 76% accuracy.