Feature Fusion Model for Efficient Skin Cancer Classification

Abstract—Skin cancer ranks as the most prevalent cancer globally, occurring from generic reasons or ultraviolet radiation, which preliminary stage detection is necessary to reduce mortal- ity rates. Currently used diagnostic methods have limitations, including human error that can lead to misdiagnosis. Also, notable limitations are demonstrated in existing deep learning ap- proaches including classification accuracy, lack of generalization, and deployment challenges. An improved feature fusion model for binary skin cancer classification using VGG19 and MobileNet as the base model has been proposed. The methodology addresses the existing limitations in classification of dermoscopic images by achieving higher accuracy, increasing both true positive and true negative simultaneously, interpretability analysis for transparency, and deployment. The proposed method achieved a classification accuracy of 95.61%, outperforming both existing models and previously reported benchmarks. A user-friendly web application has been developed incorporating the fusion model, allowing interaction with the model and real time diagnosis of skin cancer.