Deep learning with SMOTE techniques for improved skin lesion classification on unbalanced data

Skin cancer has become a major public health concern around the world, with an increasing incidence in recent decades. The morphological characteristics of skin lesions are thought to be an important component of skin cancer diagnosis and early detection. Thus, with rapid advances in image classification, more emphasis has been placed on computer-aided diagnosis (CAD) of skin lesions according to their morphological features. However, small datasets or an imbalance of skin cancer datasets are the two most important issues that can hinder the success of skin cancer detection. This paper introduces a method for dealing with class imbalance and data scarcity that is based on the Synthetic Minority Oversampling Technique (SMOTE). The improved images were then used to train the Deep Learning Convolutional Neural Network (DLCNN) model. The proposed data augmentation technique is used to generate a new skin dataset for the HAM10000 dataset using dermoscopic images of seven skin lesion classes. According to the empirical results, the improved strategy proposed in this study has statistically significant effects on improving performance with respect to accuracy (85.99%), precision (90%), recall (88%), and F1-score (88%). Moreover, the proposed classification strategy It outperforms some of the techniques used to balance melanoma detection data.