An ensemble image augmentation approach to enhance granular parakeratosis dataset

Abstract

The study discusses the revolutionizing impact of deep convolutional neural network (CNN) techniques on medical image classification, particularly in identifying skin lesions. It addresses the challenge of limited datasets for granular parakeratosis (GP) and paraneoplastic pemphigus (PNP) by employing traditional and advanced ensemble data augmentation techniques. These techniques include geometric transformations, generative adversarial networks (GANs), Cutout, and keep augment. GP affects keratinization in the groin and other regions, while PNP is associated with malignancies. The study’s relevance is enhanced by the shared imaging characteristics of the chosen conditions. By utilizing tools like U-net for segmentation, region props for feature extraction, and a support vector machine (SVM) 10-fold cross-validation model for classification, the study achieved impressive performance metrics, including 95% accuracy, 100% sensitivity, and 100% specificity when evaluated on the DermnetNZ skin lesion dataset. These findings underscore the effectiveness of augmentation in enhancing the precision of medical image classifiers and signify a substantial improvement over traditional method. Thus, the research showcases the critical role of data augmentation in overcoming data scarcity challenges and advances medical image analysis.