Evaluating sampling strategies of dermoscopic interest points

Among the most critical components of a computerized system for automated melanoma detection is image sampling and pooling of the extracted features. In this paper, we propose a new method for sampling and pooling based on a combination of spatial pooling and graph theory features. The performance of the new method is evaluated using a dataset of more than 1,500 images representing pigmented skin lesions of known pathology. In our comparisons, we include several methods ranging from simple and multi-scale sampling on a regular grid to more sophisticated approaches, such as blob and curvilinear structure detectors. Our results show that, despite its simplicity, simple sampling on a regular grid provides highly competitive performance, compared to the more sophisticated approaches, while multi-scale sampling yields only trivial improvements. However, the proposed method provides significant performance improvement in terms of sensitivity and area under the receiver operating characteristic curve (95% t-test), and the best performance in terms of specificity compared to all other methods explored.