Genetic Programming for Multiple Feature Construction in Skin Cancer Image Classification

Skin cancer is a common cancer worldwide, with melanoma being the most deadly form which is treatable when diagnosed at an early stage. This study develops a novel classification approach using multi-tree genetic programming (GP), which not only targets melanoma detection but is also capable of distinguishing between ten different classes of skin cancer effectively from lesion images. Selecting a suitable feature extraction method and the way different types of features are combined are important aspects to achieve performance gains. Existing approaches remain unable to effectively design a way to combine various features. Moreover, they have not used multi-channel multi-resolution spatial/frequency information for effective feature construction. In this work, wavelet-based texture features from multiple color channels are employed which preserve all the local, global, color and texture information concurrently. Local Binary Pattern, lesion color variation, and geometrical border shape features are also extracted from various color channels. The performance of the proposed method is evaluated using two skin image datasets and compared with an existing multi-tree GP method, ten single-tree GP methods, and six commonly used classification algorithms. The results reveal the goodness of the proposed method which significantly outperformed all these classification methods and demonstrate the potential to help dermatologist in making a diagnosis in real-time situations.