Automatic Feature Learning via Genetic Programming with Flexible Filtering for Skin Cancer Image Classification

Skin cancer images frequently contain substantial noise, which poses challenges for effective feature extraction and classification. Although existing genetic programming (GP)-based methods exhibit adaptability to diverse tasks, they frequently lack dedicated mechanisms to effectively address noise. To overcome this limitation, this paper develops GPFF (genetic programming with a flexible filtering layer), a novel approach that reduces noise and enhances the extraction of meaningful and diverse feature representations. A novel program structure is proposed, incorporating a flexible filtering layer to enhance the reliability of feature extraction by effectively reducing noise. The flexible filtering layer incorporates a variety of image filtering functions, which capture critical image characteristics across multiple domains. By flexibly selecting and combining these filtering functions, GPFF maintains robustness across various datasets. Extensive experiments on four skin cancer datasets demonstrate that GPFF consistently outperforms five traditional feature extraction methods and three GP-based methods in most cases. Further analysis shows that the flexible filtering layer improves classification performance while achieving effective feature learning without significantly increasing computational costs, underscoring its practicality for skin cancer image classification tasks.