Skin Cancer Detection with Multimodal Data: A Feature Selection Approach Using Genetic Programming

Qurrat Ul Ain, Bing Xue, Harith Al-Sahaf, Mengjie Zhang

Abstract: Melanoma is the most deadly form of skin cancer and can be treated if detected at an early stage. This study develops a skin cancer image classification method using the feature selection ability of genetic programming and multi-modal skin cancer data. This study utilizes suitable feature descriptors to extract informative features that incorporate the scale, color, local, global, and texture information from the dermoscopic images, as well as effectively utilize domain knowledge to enhance the performance of binary and multiclass classification tasks. Existing approaches mainly rely on grayscale and texture information to classify skin cancer images. Designing an effective way to combine multi-channel multi-resolution spatial/frequency information has not been well explored to improve the classification performance of complex skin cancer images. To preserve all local, global, color, and texture information simultaneously, we extract Local Binary Patterns and wavelet decomposition features from multiple color channels. The proposed method is evaluated using a dermoscopic image dataset and compared to existing deep learning and GP methods. The results conclude that the proposed method outperformed the other methods in this study. With the interpretability of GP models, the proposed method highlights important domain-specific features with high discriminating ability between different types of skin cancers. This discovery validates the potential of the proposed method to improve dermatologists’ real-time diagnostic ability.