Skin Cancer Detection from Dermatoscopic Images Using Hybrid Fuzzy Ensemble Learning Model

Malignant tissue in the skin is highly harmful. As melanoma is of identical look and lacks color variation, detection of skin cancer from dermatoscopic scans is a difficult task. The raw dataset is imbalanced that has been rarely studied. This is considered in this work to improve the accuracy level. The data have been studied and balanced using Synthetic Minority Oversampling Technique (SMOTE). The classifiers are considered as fuzzy-based classifiers that are statistical in nature and can pick the specific class, so that all the oversampled may not be useful except the informative samples. A combination of both homogeneous and heterogeneous ensemble learning termed a hybrid ensemble learning model is proposed. The homogeneity is formed by using two similar models, i.e., two adaptive resonance theory mapping (ARTMAP) models. One ARTMAP is trained with the raw imbalanced dataset, whereas the other one is trained with the balanced dataset. The heterogeneity is considered using fuzzy min-max (FMM) as the third base classifier. Its learning strategy is different from ARTMAP. Finally, the classification is performed using the rule-based neuro-Fuzzy classification (NEFCLASS) model. The open-source Kaggle HAM10000 skin dermatoscopic dataset is used for the training and detection of skin cancer. The proposed model provided 98.4% classification accuracy that competes with state-of-the-art models in the field of skin cancer detection. An improved form of ensemble learning is proved to be an efficient choice in the field of image processing.