Multi-class segmentation of skin lesions via joint dictionary learning

•An automated framework based on joint dictionary learning is proposed for multi-class segmentation of skin lesion images.•The proposed method considers data from two different feature spaces and jointly learns two dictionaries.•Final segmentation is made using a graph-cut method.•Experiments dealt with multi-class segmentation of melanoma, seborrheic keratoses and benign lesions.•Experimental results show efficiency of the proposed method even on classes having limited number of training data. Melanoma is the deadliest type of human skin cancer. However, it is curable if diagnosed in an early stage. Recently, computer aided diagnosis (CAD) systems have drawn much interests. Segmentation is a crucial step of a CAD system. There are different types of skin lesions having high similarities in terms of color, shape, size and appearance. Most available works focus on a binary segmentation. Due to the huge variety of skin lesions and high similarities between different types of lesions, multi-class segmentation is still a challenging task. Here, we propose a method based on joint dictionary learning for multi-class segmentation of dermoscopic images. The key idea is based on combining data from different feature spaces to build a more informative structure. We consider training data from two different spaces. Then, two dictionaries are jointly learned using the K-SVD algorithm. The final segmentation is accomplished by a graph-cut method based on both the topological information of lesions and the learned dictionaries. We evaluate our proposed method on the ISIC 2107 dataset to segment three classes of lesions. Our method achieves better results, specially for challenging skin lesions, compared to the only available method for multi-class segmentation of dermoscopic images. We also evaluate the performance of our method for binary segmentation and lesion diagnosis and compared the results with the other state-of-the-art methods. Experimental results show the efficiency and effectiveness of the proposed method in producing results that are more reliable for clinical applications, even using limited amount of training data.