1966 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 52, NO. 11, NOVEMBER 2005 Communications______________________________________________________________________ Counting Moles Automatically From Back Images Tim K. Lee*, M. Stella Atkins, Michael A. King, Savio Lau, and David I. McLean Abstract—Density of moles is a strong predictor of malignant melanoma, therefore, enumeration of moles is often an integral part of many studies that look at malignant melanoma. An automatic method of segmenting and counting moles would help standardize studies, compared with manual counting. We have developed an unsupervised algorithm for segmenting and counting moles from two-dimensional color images of the back torso region, as part of a study to evaluate the effectiveness of sunscreen. The method consists of a new variant of mean shift filtering that forms clusters in the image and removes noise, a region growing procedure to select can- didates, and a rule-based classifier to identify moles. When this algorithm was compared to an assessment by an expert dermatologist, the algorithm showed a sensitivity rate of 91% and diagnostic accuracy of 90% on the test set, for moles larger than 1.5 mm in diameter. Index Terms—Adaptive mean shift filters, biomedical image processing, image segmentation, moles, nevi, noise removal. I. INTRODUCTION Cutaneous malignant melanoma is a potentially lethal form of skin cancer. The mechanism of melanoma development is not yet deter- mined, but mole density has been reported as the strongest risk factor [1] with about 50% of melanoma originating from pre-existing moles [2], and moles have been considered as the precursor for the disease [3]. The important relationship between moles and melanomas makes mole counting an integral part of many melanoma studies. However, manual counting is costly and subjective and can be inconsistent, depending upon the training of the counter. An automatic method of segmenting and counting moles would help standardize studies, and would aid in the registration and tracking of moles of patients who are at a high risk of developing the disease and who need regular mole examinations. In related work, the most common mole segmentation methods are based on thresholding skin images [4]–[6]. Other methods require two steps. Step one uses radial lines or profiles [7]–[10] crossing the mole interior into the surrounding skin to determine the threshold, then in step two the true boundary position on the radial lines or profiles is de- termined. Texture features have also been exploited. Markov random field model parameters were estimated from the first principal com- ponent image in [11], while [12] used co-occurrence matrices in con- junction with a pyramid-based region growing method on the inten- sity image. Another region growing algorithm reported in [11] utilized a nonlinear diffusion technique termed “stabilized inverse diffusion equation”. Neural networks have been used to classify extracted mole features in [13]. Manuscript received June 7, 2004; revised February 13, 2005. This work was supported in part by the Natural Science and Engineering Research Council of Canada (NSERC). Asterisk indicates corresponding author. *T. K. Lee is with the School of Computing Science, Simon Fraser University, Burnaby, BC V5A 1S6, Canada. He is also with the Cancer Control Research, British Columbia Cancer Agency, 600 W 10th Ave., Vancouver, BC V5Z 4E6, Canada, (e-mail: tlee@bccancer.bc.ca). M. S. Atkins and M. A. King are with the School of Computing Science, Simon Fraser University, Burnaby, BC V5A 1S6, Canada. S. Lau is with the School of Engineering Science, Simon Fraser University, Burnaby, BC V5A 1S6, Canada. D. I. McLean is with BC Cancer Agency and the University of British Co- lumbia, Vancouver, BC V5Z 4E6, Canada. Digital Object Identifier 10.1109/TBME.2005.856301 The above methods have achieved various degrees of success. How- ever, these methods were designed to work with images with exactly one mole in the middle of the image surrounded by skin background, and these methods cannot be applied without modifications to our study, which attempts to identify all moles in a back torso. Currently, only a few methods have been described to extract multiple moles from images of large anatomic sites [14], [15]. In [14], the moles in the front and the back torso are detected by thresholding the output of a Sobel operation, which highlights the border of the moles. However, a global threshold value is difficult, if not impossible, to obtain. In [15], a stereoscopic mole mapping system has been proposed, which utilizes a pair of stereoscopic cameras and a high resolution texture camera. This system allows full three-dimensional reconstruction of the trunk using eight views when a patient is rotated on a motorized turntable, but the technical details of locating the moles are not given. Our research team has been investigating the mole counting problem and earlier results have been presented in [16], [17]. In this paper, we describe an automatic method of segmenting and counting moles, using a new unsupervised algorithm for segmenting and counting moles from two-dimensional color images of the back torso region. This will help standardize studies, and can be used to track moles and mole development. The remainder of the paper is structured as follows. The materials and methods are given in Section II, and the results in Section III. Section IV contains a discussion of the results, and the conclusions presented in Section V. II. MATERIALS AND METHODS A. Materials Data was obtained from color slides taken from an epidemiologic study on broad-spectrum sunscreen use and mole development [18]. The study photographed the back of participants during their mole ex- aminations according to the protocol specified in [19], and moles were manually counted according to the protocol specifications, which de- fine countable moles as brown to black spots that are well defined and darker than surrounding skin. The images were digitized at 2000 dpi with 24-bit color so that the resolution was about 4 pixels/mm. A test set of eight digitized images was chosen at random and labeled by a dermatologist trained for mole recognition. One of the test set of im- ages was also digitized at 48-bit color, using the same digitizer. B. Methods: Finding the Moles The first stage is to identify the back, the region of interest, from the digitized images. 1 The second stage, to find moles from skin im- ages, consists of three major steps, detailed below. The first step uses a version of the mean shift algorithm [20]–[22] that removes noise from the image while preserving the moles. The second step is a simple merging algorithm that creates large clusters of pixels and identifies the possible moles in the image. The final step classifies each candi- date as moles or not. 1) Noise Removal Using the Mean Shift Algorithm: The general mean shift algorithm [20] is a good clustering algorithm that preserves edge boundaries as well keeping a representative color or intensity of each cluster. It has been used with extensions [21], [22] for image segmentation. We built another extension with an adaptive kernel for use in segmenting moles, exploiting the fact that moles are clusters of pixels that are darker than the surrounding skin. Our version of the 1 An automatic segmentation program of the back torso was developed, but it is not used in this paper. 0018-9294/$20.00 © 2005 IEEE