Exploring the strengths of Pre-trained CNN Models with Machine Learning Techniques for Skin Cancer Diagnosis

Most people exposed to UV radiation get skin cancer, which is the most common health risk. The condition is identified when the epidermal tissues develop in an abnormal way and take on a color that is not typical for them. Strong exposure to ultraviolet light is the causative agent of melanoma. If detected at an early stage, melanoma can be cured. Cancer has the potential to spread to other areas of the body if it is not diagnosed, making it difficult to treat and eventually fatal. Visual inspection and biopsy are the traditional methods for detecting melanoma; however, the accuracy of these methods is not always reliable. Because of the risks involved, distinguishing between benign and malignant forms of melanoma has emerged as a significant research problem in the field of medical imaging. In this paper, we investigate the strengths of the CNN’s features for skin cancer diagnosis. For this, we analyze features of different CNN models like Alex-Net, Googlenet, VGGNet16, VGGNet19, ResNet18, ResNet50, ResNet101, DenseNet121 and DenseNet161. Subsequently, various machine learning methods such as Decision Tree, K-Nearest Neighbor (KNN), Logistic Regression, Naive Bayes, Random Forest, and support vector machine (SVM) were utilized to classify skin lesion. The ISIC-2016 dataset is selected to perform the experiment. The result shows that the highest accuracy is achieved with Alex-net and Densenet121 with the Random Forest classifier. Also, the results show that the fusion of multi CNN features further enhances the accuracy.