Transfer Learning and Supervised Machine Learning Approach for Detection of Skin Cancer: Performance Analysis and Comparison

Abstract

Objective: Skin cancer is one of the most lethal kinds of cancer caused by broken Deoxyribonucleic acid (DNA). This defective DNA causes cells to multiply indiscriminately and is already expanding quickly. Some research on computerized analyses of malignancy in skin lesion images has been performed. However, it is difficult to analyze these images because of light reflections, surface, color illumination, irregular lesion shapes, sizes, etc.

Method: The purpose of this research is to leverage the advantages of Transfer Learning (TL) and Supervised machine learning methods to introduce skin cancer detection in the context of imbalanced families without requiring complex feature extraction or data augmentation processes. The proposed architecture consists of two distinct CNN models: DenseNet-201, EfficientNet-B0, and six machine learning algorithms: Random Forest, SVM, XGBoost, LGBM, AdaBoost, Bagging. On an ISIC-2020 imbalanced dataset, the detection and classification performance was evaluated using various thorough assessment criteria.

Results: In this study found that XGBoost is outperforming than other supervised machine learning algorithms in terms of Accuracy (98%), f1-score (97%), Kappa (86%), and Mathew’s Correlation coefficient (79%).

Conclusion: Thus, we found that the gradient boosting machine-like XGBoost and LGBM is better than other machine learning algorithms such as SVM, Random Forest, AdaBoost, Bagging.