Enhancing Skin Cancer Detection Using Curriculum Learning In Ensemble Deep Learning Context

Abstract

Accurate classification of skin lesions, particularly for melanoma detection, remains a critical challenge in medical image analysis. Leveraging recent advances in deep learning, this paper investigates the use of curriculum learning in ensemble deep learning context for melanoma classification. To validate the proposition, three primary strategies are compared: transfer learning of CNNs using VGG16, ResNet50, and EfficientNetB0 models; ensemble learning techniques such as bagging; and curriculum learning that progressively guides training in increasing order of complexity. Experiments conducted on the ISIC 2019 and 2020 dermoscopic image datasets demonstrate that curriculum learning applied to EfficientNetB0 achieves superior classification performance, reaching an F1- score of 90.77%, outperforming conventional fine-tuning and ensemble approaches. These results underscore the potential of integrating curriculum learning in nsemble learning context with state-of-the-art CNN architectures to improve the robustness and accuracy of automated melanoma diagnosis.