A CNN Based Ensemble Deep Learning Approach for Melanoma Detection

Melanoma, the most aggressive form of skin cancer, necessitates early detection for successful treatment. This paper proposes an ensemble Convolutional Neural Networks (CNNs) approach to classify dermoscopic images of Melanoma (MEL) and Melanocytic Nevi (NV), by systematically selecting and comparing various pre-trained models with different optimizers and preprocessors through trial and error, as well as optimizing loss functions. The ensemble method takes predictions from multiple pre-trained models and selects the most reliable outcome using a custom metric. The result presents an optimal CNN ensemble model built of DenseNet121 and VGG16 with cropping and scaling, contrast enhancing, and exposure adjusting as image preprocessors that achieved a notable accuracy of 97.5% on a customized and balanced version of the Kaggle HAM10000 dataset. This approach aims to improve patient outcomes by providing a valuable tool for early melanoma detection.