Deep Learning for Skin Cancer Classification: Leveraging Feature Extraction and Transfer Learning Strategies

In this study, two deep learning classifiers examine whether there is a difference between direct classification of pre-trained ResNet model versus feature extraction before classification. Unlike seeking the opinion of a dermatologist, which can be time consuming and costly, this study augments the role of the medical diagnostics with the use of Artificial Intelligence (AI) in dermatology. Applying deep learning approaches, this study improves the quality and preciseness of skin lesion identification, which can be used in the early diagnosis of deadly diseases. Further, this study has employed two main methodologies: (1) classification of direct manner using a ResNet model without feature extraction, and (2) feature extraction utilizing a ResNet50 model after the classification of them. The method was appraised by the way it allowed to recognize the skin lesions among nine distinct categories like actinic keratosis, melanophore, and squamous cell carcinoma etc. The research consequently created 3 main findings where the direct classification method attained the final training accuracy of $95.36 \%$ and the validation accuracy of $84.11 \%$, while the feature extraction approach got similar results. Nevertheless, a closer examination of the implications of the two methods was also done and this has far-reaching significance for the clinical practice. In light of this study, it can be clearly seen that the exploration of utilizing Deep Learning (DL) techniques for skin lesion classification is crucially important, and more consistent approach that will improve diagnostic accuracy in dermatology is what we seek by utilizing AI technology.