Enhanced Skin Cancer Classification using Deep Learning and Nature-based Feature Optimization

This paper presents a skin cancer classification model that combines a pre-trained Convolutional Neural Network (CNN) with a nature-inspired feature optimization algorithm. A custom dataset comprising both malignant and benign skin cancer microscopic illustrations is derived from the ISIC dataset of dermoscopic images. Several preprocessing steps are performed on the input pictures, such as histogram equalization, gamma correction, and white balance adjustment, to improve visibility, quality, and make color corrections. Deep feature extraction and pattern recognition are conducted on both enhanced and original dataset images using the pre-trained CNN model EfficientNetB0. As a result of fusing these features, the model can capture rich details from both dataset versions at the same time. Ant Colony Optimization (ACO), a nature-inspired feature selection algorithm is applied to perform model optimization by keeping the most relevant features and discarding the unnecessary ones. The optimized feature vector is then used with various SVM classifier kernels for the skin cancer classification task. The maximum achieved accuracy of the proposed model exceeded 98% through CB-SVM while maintaining an excellent prediction speed and reduced training time.