Alleviation of Health Data Poverty for Skin Lesions Using ACGAN: Systematic Review

Abstract

Skin-based infections are one of the primary causes of the global disease burden. Digital Health Technologies powered by data science models have the potential to revolutionize global health care. Health data poverty refers to the failure of individual people, teams, or communities to profit in research or development owing to a deficiency of representative data. Generative Adversarial Network-based synthetic images can be viable solutions to health data poverty since timely detection and frequent monitoring are extremely critical for the survival of the patients. This study aims to investigate the possibility of obtaining photo - realistic dermatoscopic images of Skin Lesions via Generative Adversarial Networks (GAN), followed by distributing the images to augment the existing dataset to further enhance the performance of a Convolutional Neural Network for the task of classification. The medical and technological publications in six databases: PubMed, Web of Science, IEEE Xplore, Science Direct, Scopus, and Google Scholar were investigated. A Deep Learning pipeline has been created and a set of deep learning models such as VGG16 (Visual Geometry Group 16), DenseNet, Xception, and Inception-ResNet v2 have been assembled. We have used condition-based generative adversarial networks (GANs) besides the traditional data augmentation approaches such as rotation and scaling. To highlight the image features that eventually lead to classification are highlighted using a Local Interpretable Model-Agnostic Explanation (LIME) strategy. It was inferred from the results of the classification that DenseNet-201 with GAN Augmentation was the best individual model, with an accuracy of around 82%, while models such as VGG-16 and SVM (Support Vector Machine) were unable to compete. It was also observed that starting with the pre-trained ImageNet weights sped up the convergence and prevented models from over fitting in the absence of the regularization effect of augmented data. However, the exploitation of the data was still not perfectly optimal, as over fitting with data augmentation and early stopping was observed, which can be used by more extensive data augmentation techniques. The GAN augmentation showed to reduce the data imbalance and increase the data percentage of the less representative classes. A data augmentation approach based on synthetic data that has been obtained from GAN helps us to classify images of lesions of the skin with high accuracy. We can also infer from the results obtained that, enriching the data with GAN-produced data samples results in a significant performance increase. In the field of medical imaging, where particularly large training datasets are not available, novel data augmentation and generation procedures can be beneficial.

Keywords

• Deep learning
• health data poverty
• machine learning
• scalability
• digital health
• GAN