Automatic segmentation of dermoscopic images by iterative classification

Abstract

Accurate detection of the borders of skin lesions is a vital first step for computer aided diagnostic systems. This paper presents a novel automatic approach to segmentation of skin lesions that is particularly suitable for analysis of dermoscopic images. Assumptions about the image acquisition, in particular, the approximate location and color, are used to derive an automatic rule to select small seed regions, likely to correspond to samples of skin and the lesion of interest. The seed regions are used as initial training samples, and the lesion segmentation problem is treated as binary classification problem. An iterative hybrid classification strategy, based on a weighted combination of estimated posteriors of a linear and quadratic classifier, is used to update both the automatically selected training samples and the segmentation, increasing reliability and final accuracy, especially for those challenging images, where the contrast between the background skin and lesion is low.

Figure 1

A flowchart of the proposed ICS algorithm showing the main processing modules.

Figure 2

(Left) The small automatically selected green boxes correspond to the skin seed regions used as initial training. The location is based on the brightest luminance in each quadrant. The blue box corresponds to the seed region for the lesion, selected as the region statistically most different from the skin seeds. The three middle images show intermediate iterative steps at iterations j = {1,2} and the final iteration which is j = 7 in (a), j = 3 in (b). In these maps, yellow and blue pixels are those classified with high confidence as skin and lesion, respectively. The maps include also red and green pixels, corresponding to those pixels classified with low confidence as skin and lesion, respectively. The white contour in the rightmost figure is the final border obtained after postprocessing.

Figure 3

Average accuracy (η) of the detected seed regions for the images in the test set, according to the diagnosis. For visualization purposes, instead of the number of lesions, the vertical axis shows the corresponding percentage out of 100 benign and 22 malignant lesions.

Figure 4

Dispersion of the accuracy of segmentation, quantified by the four distinct border measures, showing the performance of the proposed ICS method against the different segmentation methods analyzed. D1, D2, and D3 corresponds to the manual segmentations provided by dermatologists. SRM, AT, and ICS are automatic methods. LDA, QDA, and SVM are supervised methods that here include only for reference purposes. The accuracy scores are computed using the mask generated by majority voting of the manual segmentations provided by three dermatologists as ground-truth reference.

Figure 5

Average values of the accuracy of segmentation, quantified by the four distinct border measures, showing the performance of the proposed ICS method against the different segmentation methods analyzed. The lesions are grouped in three disjoint sets: low, intermediate, and high contrast between skin and lesion. The mask generate by majority voting of the manual segmentation by three dermatologists is used as ground-truth reference. The bars refer to low (orange), medium (blue), and high (pink) contrast lesions.

Figure 6

Average values of the accuracy of segmentations, quantified by the four distinct border measures, showing the performance of the proposed ICS method against the different segmentation methods analyzed. The lesions are grouped by histopathological diagnosis. The mask generate by majority voting of the manual segmentation of the three dermatologists is used as reference. The bars refer to benign (blue) and malignant (green) lesions.

Figure 7

(a) Number of iterations for convergence and the (b) frequencies of the automatically selected value for the weight λ showing how the LDA (λ = 0) and QDA (λ = 1) posteriors where combined according to (1) in the proposed ICS algorithm. Results are summarized according to three intervals of possible λ values.

Figure 8

Example of lesions in the test set and the border provided by (a) the three dermatologists and (b) the automatic borders by AT (black border), SRM (gray border), and the proposed ICS algorithm (white border). None of the methods performed better in all cases.

Figure 9

Example of lesions in the test set and the border provided by (a) the three dermatologists and (b) the automatic borders by AT (black border), SRM (gray border), and the proposed ICS algorithm (white border). None of the methods performed better in all cases.