2.2.2. Correlation-Based Approach

This method assumes that there are a finite number of mole patterns that dominate in the whole-body images. Thus, proper identification of such patterns allows detection of the majority of skin moles based, e.g., on a correlation method. Masks that contain most typical mole patterns (selected based on analysis on acquired images) move across the image to estimate a set of correlation functions. The normalized cross-correlation between the image and the template is estimated by the formula:

where

f(x,y)—is the image,

t(x,y)—is the lesion template,

t¯— is the mean of template,

f¯u,v—is the mean of f(x,y) in the region under the template.

Next, spots are detected in the coordinates of the correlation matrix where it reaches local maxima. A set of discriminated spots is a combination of partial results separately obtained for correlation with each mask.

These spots that contribute to the pattern set were validated by a medical doctor as “for inspection”. The choice of the proper spot masks is the essential factor of the developed algorithm. Due to this, several masks with the patient’s spots were extracted from the images and the cross-correlation was calculated. The masks that lead to the high value of correlation maxima, and to a large number of maxima should be considered as mole patterns. Such masks are characterized with the biggest similarity to the greatest number of other spots. Thus, they are good candidates for finding similar types of spots in other images (including those that are not a part of currently analyzed set). An exemplary set of masks is shown in the Figure 5.

Exemplary masks with the spots.

The correlation matrix is generated for every mask and the processed image. Searching the spots is based on finding the local maxima values in the correlation matrix (shown in Figure 6a). A few steps must be performed to delete unnecessary artefacts (Figure 6b). First, the areas where the correlation is less than zero are deleted (Figure 6c). Second, a binary image is created by thresholding the correlation matrix. Finally, searching for spots that are bigger than 1 mm and smaller (or equal) than the biggest mask image takes place. The rest are rejected (Figure 6d). The combination of all spots detected by each mask generates the final image with detected spots, as shown in Figure 6e.

Subsequent steps of spot detection (one mask analysis). (a) correlation matrix, (b) suppressed local maxima’s below threshold, (c) artefacts removed, (d) spots detected for given mask,(e) combined results for all masks.

Finally, detected spots are segmented using the Active Contour (AC) method [21] using the Y channel of YCbCr color space. Centers of the masks are the starting points for the AC function for region expansion. Sample segmentation results are presented in Figure 7. The method’s operation is summarized by Algorithm 2.

Exemplary masks 1 and its segmentation 2—by dermatologist, 3—by the software.