Skin Cancer Detection: Leveraging Hybrid Deep Learning Models and Traditional Machine Learning Classifiers

Skin cancer is a significant worldwide health concern, necessitating efficient observation and treatment methods. We explore the potential of deep neural network techniques to automate skin cancer identification, addressing challenges like classification complexity and dataset scarcity. The research makes use of two publicly accessible datasets of skin cancer photographs, featuring 3311 and 5000 photos, respectively. Our approach combines established deep learning models like ResNet-50, VGG16, MobileNet, ShuffleNet and DenseNet-201 with novel hybrid structures to leverage their complementary strengths. The methodology initiated with data pre-processing, encompassing steps such as data splitting, normalization, reshaping, and encoding. Augmentation techniques were subsequently applied to augment the dataset volume. Following pre-processing, traditional deep learning models were independently evaluated from scratch. Subsequently, these models were re-evaluated in conjunction with machine learning classifiers. Hybrid models were then constructed from the individual models and subjected to assessment. Additionally, hybrid models paired with Support Vector Machine (SVM) classifiers were assessed. Throughout the experimentation, the Adam optimizer consistently demonstrated effective performance. Notably, SVM classifiers exhibited superior performance among the classifiers measured. Accuracy-Loss plots, Confusion Matrix, and Receiver Operating Characteristic (ROC) curves were used for a comprehensive analysis. The suggested hybrid model of VGG16 and ResNet-50 with SVM classifier outperformed the conventional DL models and ML classifiers during the anatomizing process with consistent and dependable performance. Accordingly, the suggested model demonstrated accuracy on the corresponding datasets of 90.64% and 93.61%, indicating its capability as a tool aimed at skin cancer early detection