Geometric Transformations vs Noise Induction: Comparison of data augmentation techniques for dermoscopic image analysis Transformaciones geométricas vs Inducción de ruido: Comparación de técnicas de aumentado de datos para análisis de imágenes dermoscópicas

Abstract

The HAM10000 dataset, a collection of dermatoscopic images of skin lesions, has become a valuable resource for research in dermatology and machine learning. This study focuses on evaluating the efficiency of two data augmentation techniques applied to images from the HAM10000 skin cancer dataset.  The techniques evaluated in this context were geometric transformations and Gaussian noise induction. In the methodological phase, the Principal Component Analysis (PCA) technique was implemented to compare the original images with those augmented by each approach. This analysis allowed a deeper understanding of the modifications introduced by each technique, offering insights on the preservation of relevant features for skin lesion classification. The results obtained revealed superior performance when employing the Gaussian noise induction technique. This technique proved to be particularly effective in improving the quality of the data set, contributing positively to skin cancer diagnostic tasks. The analysis through PCA not only supported the efficacy of the Gaussian noise induction technique, but also provided detailed insight into how this technique preserves crucial information during the data augmentation process. this study not only highlights the relevance of the HAM10000 dataset in dermatological research, but also emphasizes the importance of selecting appropriate data augmentation techniques, with Gaussian noise induction emerging as a highly efficient option for improving the accuracy of machine learning models applied to medical imaging in the context of skin cancer.