Employing Up-Sampling and Augmentation for Imbalanced Datasets in Skin Cancer Classification Using CNN Architecture

Skin cancer represents a pressing global health concern, characterized by its widespread incidence and substantial impact on public health. Early detection is critical to prevent lasting harm, yet current diagnostic approaches are subjective and prone to errors. Previous research has made strides in developing automated classification systems; however, they often struggle with dataset imbalances, insufficient variability in the data, and complex models that limit real-time applicability, which can hinder robust model performance and generalization. This study proposes a classification system using transfer learning with ResN et-50 for skin cancer lesions, leveraging the HAM10000 dataset, which consists of 10,015 dermatoscopic images. Up-sampling and data augmentation techniques were employed to address data imbalances, resulting in a dataset of 45,844 images with balanced class distributions. This well-structured approach yielded impressive results, including a 98% accuracy rate. These findings highlight the effectiveness of the ResN et-50 model with customized configurations in improving skin cancer lesion classification on imbalanced datasets, significantly advancing early detection efforts. Future research in this area could explore further refinements in preprocessing techniques and the application of advanced machine-learning algorithms to continue enhancing diagnostic accuracy and efficacy in clinical settings.