Performance Evaluation of Oversampling Methods on Deep Learning-Based Skin Cancer Classification

The field of dermatology faces considerable challenges when it comes to early detection of skin cancer. Our study focused on using different datasets, including original data, augmented data, and SMOTE oversampled data, to identify skin cancer. Our dataset consisted of images of skin lesions from the MNIST Skin Cancer dataset (HAM 10000), including samples of both cancerous and benign cases in the dataset. We employed data augmentation to expand the dataset’s size and increase the diversity of skin lesion features. Furthermore, to tackle class imbalance in the dataset, we applied the SMOTE oversampling technique to generate synthetic samples for the under-represented group. With the original, augmented, and SMOTE oversampled datasets, we trained a Convolutional Neural Network (CNN) model. The performance of the model was evaluated using accuracy, recall, precision, and F1-score. The comparison between the results obtained from the original data, augmented data, and SMOTE oversampling data clearly revealed distinctions in performance. Our findings clearly demonstrate that employing data augmentation and SMOTE oversampling can significantly enhance the efficacy of skin cancer detection.