Multi-classification of Skin Cancer Using Multi-model Fusion Technique

Abstract

Skin cancer is increasingly common worldwide, with melanoma and basal cell carcinomas being the primary subtypes, responsible for most of the related deaths. Therefore, screening for early identification and classification is essential for proper treatment, a task achievable through deep learning techniques. While numerous studies have addressed this problem domain, challenges such as limited accuracy, deployment on edge devices, computational costs, execution times, and manual feature extraction procedures persist. In this research work, a novel deep learning architecture is proposed to address these issues. Three pre-trained deep learning architectures are utilized namely VGG-16, VGG-19, and ResNet-50, through transfer learning to develop a composite model named "Multi-Model Fusion for Skin Cancer Detection (MMF-SCD)". A composite feature vector is generated by these CNN models and passed through a final dense layer with a SoftMax activation function for skin cancer classification. Adam optimization algorithm is applied. TensorFlow and Keras libraries are employed to develop MMF-SCD model. Rectified linear unit is applied as an activation function for the output of convolutional layers of MMF-SCD. This model is developed on Jupyter notebook using open-source web platform namely google Collaboratory in python programming language. The dataset, retrieved from public dataset repository, consists of seven of the most common skin cancer diseases: actinic keratosis, basal cell carcinoma, dermatofibroma, melanoma, nevus, pigmented benign keratosis, and vascular lesion. MMF-SCD has demonstrated significantly improved accuracy compared to previous studies, achieving an accuracy gain of 97.6%.