Effective skin cancer classification by modified and optimized inception-ResNet-V2 model

Artificial Intelligence tools are flourishing in biomedical diagnosis, particularly in oncology. The prediction of skin cancer from dermoscopic images using deep learning neural networks has gained importance in recent years because of their inherent early non-invasive diagnostic capabilities. This study presents the results of the classification of benign nevus and malignant melanoma lesions using a deep learning model. The model incorporates an efficient pre-processing stage powered by median filtering and a class-balancing stage powered by the Synthetic Minority Oversampling Technique (SMOTE) to improve classification results. First, the classification efficiency of four pre-trained models, namely, ResNet50, EfficientNet B0, Inception-V3, and Inception-ResNet-V2, were studied, and the results revealed that they achieved accuracies of 93.90%, 94.37%, 94.87%, and 95.77%, respectively. Second, the effect of optimization and hyperparameter tuning on the Inception-ResNet-V2 model is studied considering Adam, Nadam, and AdaMax optimizers, with fivefold cross validation. The experimental results revealed that the AdaMax optimizer with validation achieved an overall consistent performance with accuracy, sensitivity, and specificity of 97.65%, 96.67%, and 98.92%, respectively. The results support the efficacy of the model in predicting skin cancer malignancy; thus, its integration into clinical practice could benefit healthcare services.