Dysplastic nevus. A, Dermoscopic view (×16). B, Evaluation of geometric variables, the difference between the lesion and the circle of equal area, minimum and maximun diameters, and best symmetry axis are shown (area = 31 mm2, perimeter = 40 mm, variance of contour symmetry = 0.76%, minimum diameter = 5.3 mm, maximum diameter = 7.2 mm, circularity = 0.82%). C, Evaluation of islands of color: white, background areas' imbalance = 0.03%; dark gray, dark areas = 0.13%; and magenta, transition areas = 0.19%. Note the absence of blue-gray dominant areas and blue-gray dominant areas' imbalance.
Melanoma in situ. A, Dermoscopic view (×16). B, Evaluation of geometric variables, the difference between lesion and the circle of equal area, minimum and maximun diameters, and best symmetry axis are shown (area = 20 mm2, perimeter = 24 mm, variance of contour symmetry = 0.91%, minimum diameter = 3.9 mm, maximum diameter = 6.6 mm, circularity = 0.72%). C, Evaluation of islands of color: white, background areas' imbalance = 0.15%; dark gray, dark areas = 0.12%; magenta, transition areas = 0.11%; and blue, blue-gray dominant areas = 0.10% and blue-gray dominant areas' imbalance = 0.06%.
Customize your JAMA Network experience by selecting one or more topics from the list below.
Andreassi L, Perotti R, Rubegni P, et al. Digital Dermoscopy Analysis for the Differentiation of Atypical Nevi and Early Melanoma: A New Quantitative Semiology. Arch Dermatol. 1999;135(12):1459–1465. doi:10.1001/archderm.135.12.1459
Copyright 1999 American Medical Association. All Rights Reserved. Applicable FARS/DFARS Restrictions Apply to Government Use.1999
To use a digital dermoscopy analyzer with a series of "borderline" pigmentary skin lesions (ie, clinically atypical nevi and early melanoma) to find correlation between the studied variables and to determine their discriminating power with respect to histological diagnosis.
A total of 147 pigmentary skin lesions were histologically examined by 3 experienced dermatopathologists and identified as nevi (n = 90) and melanomas (n = 57). The system evaluated 36 variables to be studied as possible discriminant variables, grouped into 4 categories: geometries, colors, textures, and islands of color.
University medical department.
A sample of patients with excised pigmentary skin lesions (nevi and melanomas).
Main Outcome Measures
Sensitivity, specificity, and accuracy of the model for evaluating "borderline" pigmentary skin lesions.
After multivariate stepwise discriminant analysis, only 13 variables were selected to compute the canonical discriminant function.
The present method made it possible to determine which objective variables are important for distinguishing atypical benign pigmentary skin lesions and early melanoma.
BECAUSE advanced cutaneous melanoma is still incurable, early detection, by means of accurate screening, is an important step toward a reduction in mortality.1-6 In most cases, the screening tests are simple and only require a brief skin examination by a trained professional.7 Of the methods for evaluating pigmented skin lesions (PSL), the ABCD approach is one of the most widely used.7,8 It is based on the study of color and a few coarse geometric variables. In some studies, it has been found to have high sensitivity (82%-96%) at the expense of specificity (28%-46%).9 However, for the many difficult or borderline lesions, diagnostic accuracy is only slightly above 60% even in specialized centers.4,10
Noninvasive diagnostic methods such as dermoscopy or epiluminescence light microscopy (ELM) have been developed in an attempt to improve diagnostic accuracy, especially in cases of borderline PSL. Used by Weiss in 1917, Saphier in 1921, and Hinselmann for detailed examination of the skin,11 ELM was later used for the diagnosis of PSL by Goldman and Younker12 and MacKie.13 It allows in vivo inspection of the lesional surface by means of a microscope. With the addition of the oil immersion technique, the epidermis becomes translucent, permitting macroscopic evaluation of the dermo-epidermal junction. Many studies have shown that this method improves diagnostic accuracy by 20% to 30% with respect to simple clinical observation, especially by an expert dermatologist.14-17 It is most valuable in the differential diagnosis of equivocal skin lesions and can be used for pattern analysis of PSL. Many studies have furnished specific epiluminescence diagnostic criteria, paving the way for a new semiology that provides the study of patterns, colors, intensities of pigmentation, configuration, regularity, and margin and surface characteristics.12-19 However, qualitative evaluation of the many morphologic characteristics of PSL observable by ELM is often extremely complex and subjective.20,21 With the aim of obviating these problems of qualitative interpretation, methods based on mathematical analysis of PSL have recently been developed.22,23 These methods are based on computerized analysis of digital images obtained by ELM.24-34
In the present study we used a digital dermoscopy analyzer (DDA) with a series of "clinically difficult" PSL, both malignant and benign in nature. We evaluated a large set of morphologic variables. Stepwise linear discriminant analysis was then used to identify an optimal subset of features for classification purposes, to find correlation between variables, and to determine their discriminating power with respect to histological diagnosis.
Between 1991 and 1998, 4650 PSL were observed in the Department of Dermatology of Siena University, Siena, Italy. Many of them were excised (n = 1220); they included 297 melanomas and 923 benign PSL. Their images were digitalized and saved in a computer archive. From this archive, 147 PSL were selected. As inclusion criteria we considered clinically atypical flat impalpable lesions, 0.4 to 1.2 cm in diameter, belonging to different subjects. All the lesions were of difficult diagnostic interpretation and were therefore suitable for the purpose of morphologic and parametric evaluation of early melanoma. Cases of familial cutaneous malignant melanoma, acral lentiginous melanoma, and malignant lentigo melanoma were excluded. The PSL were examined by 3 experienced dermatopathologists and identified as nevi (n = 90) and melanomas (n = 57).
The lesions were obtained by oil epiluminescence microscopy (DDA-Mips) at a magnification of ×16. The patient lay on the examination table and the skin around the lesion was arranged orthogonal to the incident light. If the lesion and/or surrounding skin was hairy, the hairs were carefully removed with scissors or a razor. After 5 minutes of acclimatization to room temperature (25°C), the lesion was recorded as a digital signal and stored. The lesions were then surgically removed and histological examination was performed. All digital images were analyzed with appropriate algorithms.
The equipment consisted of a surface microscope with a 150-W light source at 3200°K, which provides 5 magnifications from ×6 to ×40, allowing a field of view horizontally ranging from 6 mm to 4 cm, connected with a 3-CCD (charge-coupled device) video camera. The camera was calibrated weekly using special Kodak paper for white balance. The components of the video signal were connected to a frame-grabber mounted in a computer equipped with a 1.2-gigabyte hard disk and a magneto-optical drive for image storage. The system (DDA-Mips) runs under Microsoft Windows (Microsoft Corp, Redmond, Wash), and all the software was written in language C/C++ (M.B. and G.D.'E.).
The choice of the most useful features to extract from digital images depends on the results of epiluminescence pattern analysis. Although the system saves the microscope magnifications along with the texture analysis, offering an error-free objective evaluation, the different magnifications could confuse clinicians wanting to make subjective comparisons of lesions. In this article we only discuss the ×16 images. The system uses an algorithm based on the Laplacian-of-Gaussian for the segmentation procedure and a zero-crossing algorithm for the border automatic outline.35 Once the borders had been automatically detected and traced, the system evaluated 36 variables to be studied as possible discriminant variables. The reproducibility was first tested on digitalized images of 100 lesions belonging to 20 subjects (1 PSL for each patient recorded 5 times at regular 15-minute intervals). Absolute differences between single measurements and mean values of a given lesion or variable never exceeded 3% of the mean value. The studied variables belonged to 4 categories: geometries, colors, textures, and islands of color (ie, color clusters inside the lesion). Geometries includes area, perimeter, maximum and minimum diameters, variance of contour symmetry, circularity, and fractality of borders. Colors includes red, green, and blue lesion mean values; red, green, and blue healthy skin mean values; red, green, and blue lesion deciles; red, green and blue lesion quartiles, and skin-lesion gradient mean value. Texture includes mean contrast and entropy of the lesion and contrast and entropy fractality. Islands of color includes peripheral dark regions, dark area, dark regions imbalance, imbalance, green area, green dominant regions imbalance, blue-gray area, blue-gray regions, transition area, transition regions imbalance. The meanings of all variables are explained in Table 1, Table 2, Table 3, and Table 4.
Stepwise discriminant analysis was applied to the whole set of variables.36 This multivariate method identifies the best group separability of a selected subset of variables. Forward and backward selection of variables was performed; at each step the variable that added the most (the least) separation of the groups was included in (excluded from) the analysis. In this way, the set of variables giving the best linear discrimination was found. The variables to include or exclude were chosen on the basis of Wilks lambda (λ), which takes values from 0 to 1. Small (large) values indicate good (poor) separation between groups. The significance of the change in λ obtained by including or excluding a variable in the previous set of included variables was tested using the F statistic.37 Group distinction was evaluated using a classification matrix. Each data unit was assigned to a group by a linear classification function found by the procedure. The sensitivity and specificity were then calculated. A jackknife validation procedure was used to reduce the bias in group classification. According to this procedure, cases are eliminated one by one in the computation of classification algorithm performance,37 enabling the prediction power of the classifier to be evaluated on new data units.
The DDA-Mips system gave high-quality images acquired in real time so that it was possible to examine all the features revealed by ELM. Moreover, the digital images were subjectively almost as good as conventional photographs (Figure 1 and Figure 2). Image resolution was 45 pixels/mm at a magnification of ×16. The system was easy to use. Objective evaluations were performed automatically in real time simply by clicking a button; no modification by the operator was necessary. Graphic windows showed the objective results, which were readily understood. The operators were able to use the program without any special training.
Histological examination showed that all melanomas (n = 57) had a thickness less than 0.5 mm (32 in situ melanomas). The benign PSL were 42 dysplastic nevi and 48 junctional and compound nevi. Histopathological diagnosis of melanoma and dysplastic nevi was made according to the criteria of the National Institutes of Health Consensus Conference.4 Histopathological diagnosis showed a discordance of about 3% (3 dysplastic nevi and 2 in situ melanomas). Such cases were classified as melanoma or nevi when at least 2 of 3 pathologists agreed on the diagnosis.
Table 1, Table 2, Table 3, and Table 4 show the descriptive statistics of all the variables considered. Table 5 shows the results of univariate discriminant analysis (step 0 of the stepwise procedure). The 36 variables are listed in descending order of separation power, ie, in order of decreasing values of F and increasing values of λ. Twenty-one variables reached a statistical significance of 95% (P<.05; F>3.84). However, at the end of the multivariate stepwise discriminant analysis, only 13 variables were selected. Table 6 lists these 13 variables in the order in which they were included. The other 23 variables did not reach the inclusion threshold. The list of variables ordered at the last step of the Wilks λ and F procedures was slightly different from that obtained by univariate analysis, due to correlation among some discriminant variables. For example, after inclusion of the most discriminating variable, peripheral dark regions, at the first step, imbalance of dark regions completely lost all discriminating power on inclusion, despite the fact that it was the second most powerful discriminant in the univariate analysis. This means that these 2 variables have almost the same diagnostic content. The jackknife classification matrix showed good group distinction with the set of 13 variables. Of 90 cases of nevi, 79 (88%) were predicted as nevi and 11 (12%) were predicted as melanoma. Of 57 cases of melanoma, 11 (19%) were predicted as nevi and 46 (81%) were predicted as melanoma. The percentage of cases correctly classified was 85% with a sensitivity of 88% and a specificity of 81%.
In the past few years ELM has been accepted as a useful aid for the diagnosis of melanoma, especially for small lesions and those difficult to interpret.14,15 One of the problems with ELM is the lack of well-standardized classification criteria.38 Reference terminology based on an essential set of variables and descriptors was developed by the Consensus Conference in 1990,20 however these standards have since undergone numerous updates and modifications.39,40 Many variables related to ELM patterns, the objective value and reproducibility of which are dubious, are today used for the differential diagnosis of PSL.20 The normal descriptive criteria for melanoma (eg, radial streaming, pseudopods, gray-blue veil) or even benign PSL (regular pigment network) are already complex to quantify, without the further complication of ELM patterns found in both types of lesion but disposed in different ways (eg, pigment network, brown globules, black dots). In such cases, extremely subjective definitions such as "uniform size regularly arranged" or "variable size regularly arranged" or even "uniform size irregularly arranged" instead of "variable size irregularly arranged," have been used. Moreover, for some ELM features, it is not their presence or disposition that is the diagnostic indicator, but rather the extent to which they manifest in a PSL with respect to other morphologic characteristics of the lesion itself and the surrounding skin.20 Problems such as these made it necessary to use methods such as DDA by which images can be transformed into numbers.22,23,26 Then algorithms can be designed to obtain objective measurements from unambiguous diagnostic characteristics in line with the main ELM features so far recognized.41 Moreover, the DDA-Mips system makes it possible to define and investigate new variables, such as islands of color, useful for early diagnosis of PSL at risk.41
Our aim was to evaluate the discriminating power of numerical DDA variables for early melanoma and benign PSL with similar clinical and ELM features. Univariate analysis of single variables suggested that all groups (geometries, texture, colors, and islands of color) were important for describing and distinguishing benign PSL and early melanoma. However, the texture and islands-of-color variables seemed to have more weight than the other variables as also reported by Kenet et al24 (Table 5). This trend was confirmed by the multivariate analysis, in which these variables constituted 62% (8/13) of the statistically significant ones; 5 of them were in the islands of color group (peripheral dark, background regions imbalance, blue-gray area, dark area, and imbalance) (Table 6). This new algorithm (islands of color) was designed to identify clusters of homogeneous color inside the borders, according to the subjective definitions of traditional dermoscopic analysis. By relative color analysis of the lesion with respect to the surrounding healthy skin, the system identifies 6 different areas and their properties.42 Like the islands of color, the texture variables were also important in the discriminant analysis. Contrast was the most significant followed by fractality of contrast and entropy. The only nonsignificant texture variable was fractality of entropy, which turned out to be completely explained by entropy. All this shows that the microinternal disorder expressed by the islands of color and texture variables is more accentuated in early melanoma than benign PSL. Of the geometric variables, all highly significant in the univariate analysis, only perimeter and minimum diameter emerged as important in the multivariate analysis. Minimum diameter canceled out the statistical significance of all the other geometric variables, except of course perimeter. Of the many colorimetric variables, only the decile and quartile of green, which correspond to the slate green, and the quartile of blue were still significant after multivariate analysis.
Although our aim in this study was not to classify benign and malignant PSL, the correlation matrix obtained from the multivariate analysis raises other issues. In fact, with the above set of 13 significant variables, the jackknife classification matrix indicated good group distinction. The percentage of cases correctly classified was 85% with a sensitivity of 88% and a specificity of 81%, which is high considering the type of lesion used for the study. A greater sensitivity could be obtained at the expense of specificity by changing the decision threshold of the classification procedure. Moreover, a further increase in accuracy is probably possible by using sophisticated nonlinear classification procedures, such as logistic models or neural networks43; adding anamnestic and phenotypic variables, such as number of nevi,44 objectively measured skin color, phototype45; and increasing the number of lesions studied.
In conclusion, the present method made it possible to determine which objective variables are important for distinguishing atypical benign PSL and early melanoma. These results may not apply to acral lentiginous and malignant lentigo melanoma, which were excluded from the analysis. Many of the geometric variables (area, symmetry, outline, circularity, and borders), which are often the only ones that can be easily evaluated by the naked eye, are certainly less important than variables related to microinternal disorder (texture and islands of color). Such variations within a lesion's borders can also be detected by the trained eye performing ELM without a computer. However, by identifying the relative importance of various morphologic features of ELM photographs, the results of this study may be useful for training the eye to perform ELM without a computer, and perhaps to confirm a computer's ELM diagnosis. This suggests that the present method is potentially a formidable tool for research and teaching.
Accepted for publication June 14, 1999.
Reprints: Pietro Rubegni, MD, Istituto di Scienze Dermatologiche, Università degli Studi di Siena, Policlinico "Le Scotte," 53100 Siena, Italy (e-mail: Rubegni@unisi.it).
Create a personal account or sign in to: