Mining biomedical images with density-based clustering

Abstract

Density-based clustering algorithms have recently gained popularity in the data mining field due to their ability to discover arbitrary shaped clusters while preserving spatial proximity of data points. In this work we adapt a density-based clustering algorithm, DBSCAN, to a new problem domain: identification of homogenous color regions in biomedical images. Examples of specific problems of this nature include landscape segmentation of satellite imagery, object detection and, in our case, identification of significant color regions in images of skin lesions (tumors). Automated outer and inner boundary segmentation is a key step in segmentation of structures such as skin lesions, tumors of breast, bone, and brain. This step is important because the accuracy of the subsequent steps (extraction of various features, post-processing) crucially depends on the accuracy of this very first step. In this paper, we present an unsupervised approach to segmentation of pigmented skin lesion images based on DBSCAN clustering algorithm. The color regions identified by the algorithm are compared to those identified by the human subjects and the Kappa coefficient, a statistical indicator of computer-human agreement, is found to be significant.