Irregularity index: A new border irregularity measure for cutaneous melanocytic lesions
Introduction
Cutaneous melanocytic lesions are the common brown pigmented skin lesions known as moles. These lesions are formed by nests of specialized cells called melanocytes in the outer layer of the skin. Melanocytes produce a pigment called melanin, which protects the body from harmful ultraviolet radiation. When melanocytes proliferate, a skin lesion is formed and a pigmented mark appears. The majority of such pigmented skin lesions are benign, medically known as melanocytic nevi; however, some of the lesions are malignant, medically known as cutaneous malignant melanomas. Early diagnosis of malignant melanomas is crucial to treatment processes because the survival rate is inversely proportional to the thickness of the lesion (Balch et al., 1989). Due to the increasing incidence rate in the last three decades (Armstrong and English, 1996), many dermatologists are searching for non-invasive computer-aided diagnostic systems which can help health care providers to diagnose early lesions and to improve diagnostic accuracy and consistency. Furthermore, such a diagnostic device can reduce the number of skin biopsies (Stoecker and Moss, 1992, Hall et al., 1995, Fleming, 2000).
Studies have been conducted to investigate the possibility of detecting melanomas using image analysis techniques. The clinical surface view features (border irregularity, asymmetry, texture, colour and lesion size) and clinical subsurface view features (pigment networks, brown globules, black dots, radial streaming and pseudopods) have been extracted and analysed (Cascinelli et al., 1987, White et al., 1991, White et al., 1992, Schindewolf et al., 1992, Binder et al., 1994, Ercal et al., 1994, Hall et al., 1995, Seidenari et al., 1995, Gutkowicz-Krusin et al., 1997, Menzies et al., 1997, Colot et al., 1998, Fleming, 2000). These features were often extracted from colour images, but wide spectrum images including invisible wavelengths have also been investigated (Cotton et al., 1999, Bono et al., 1999, Elbaum, 2000). Based on an optical model of skin, Cotton et al. (1999) was able to derive features such as the amount of dermal blood and epidermal melanin from colour and near-infrared images. Although the above features (surface view or subsurface view, structure-based or texture-based) provide important evidence for melanoma, none of them can provide a proper diagnosis by itself. A classifier is required to analyse all features to generate a diagnosis.
Among all features, lesion shape is one of the important symptoms for diagnosing melanoma. Clinically, benign nevi are often described as small skin lesions with uniform colour. They usually have a round or oval shape border. On the other hand, malignant melanomas usually appear as enlarged nevi with multiple shades of colours, and their borders tend to be irregular and asymmetric with protrusions and indentations (Maize and Ackerman, 1987, Rivers, 1996). Fig. 1 shows a typical benign nevus (Fig. 1(a)) and a malignant melanoma (Fig. 1(b)). Among these melanoma features, border irregularity has been reported as the most significant factor in clinical diagnosis (Keefe et al., 1990). Histologically, protrusions along the lesion border may suggest abnormality and excessive cell growth in a sub-population of the melanocytes, while indentations may suggest an occurrence of regression in the melanoma. Therefore, marked border irregularities may be indicative of malignancy of a pigmented skin lesion. In this paper, we concentrate on how to measure border irregularity effectively so that the measurement can be used as an input feature by a classifier for malignant melanomas.
When a lesion border is studied carefully, we notice two types of irregularities: texture and structure irregularities (Claridge et al., 1992). Texture irregularities are the fine variations along the lesion border. Detecting and measuring texture irregularities may be subject to noise from the hardware imaging devices and/or software programs. On the contrary, structure irregularities, which are general undulations of the perimeter, may infer the abnormal histological signs discussed in the previous section, and have a higher correlation with melanomas (Claridge et al., 1992). Therefore, measuring structure irregularities accurately is important for diagnosing melanomas. Fig. 2 shows both types of irregularities with three border outlines extracted from pigmented skin lesions. Lesion border A has no structure protrusion and indentation, but a lot of texture irregularities. Lesion border B has a structure protrusion at the top of the border but has less texture irregularity than the other two borders, while lesion border C has a prominent structure protrusion and indentation at the bottom of the border.
Most previous studies used some common shape descriptors such as the compactness index (CI) (White et al., 1992, Golston et al., 1992, Ercal et al., 1994, Stoecker et al., 1995, Colot et al., 1998) or the fractal dimension (FD) (Claridge et al., 1992, Claridge et al., 1998, Hall et al., 1995, Ng and Lee, 1996) to measure border irregularity. Unfortunately, the CI is sensitive to noise along the border. Alternatively, Mandelbrot’s FD (Mandelbrot, 1982) has been used extensively to measure the roughness (jaggedness) of a border or a surface for many applications (Pentland, 1984, MacAulay, 1989, Caldwell et al., 1990, Chaudhuri and Sarkar, 1995). However, the FD does not measure structure irregularities. In an attempt to capture major structural features, Claridge et al. (1992) designed a measure called Structure Fractal Dimension (SFD), where the lesion border is smoothed slightly before the FD is computed. There are problems with this measure which we discuss later in Section 4.2.
In our earlier work, we reported a new measure for border irregularity, called Sigma-Ratio (SR) (Lee et al., 1999), which is based on the number of Gaussian smoothing iterations required for eliminating all concavities along the lesion border. We showed that this simple index is more sensitive to structure indentation and protrusion than the CI, FD and SFD. However, there are some shortcomings of the SR measure. First, the SR measure is non-linear. Second, the SR measure is sensitive to a long and narrow indentation such as the one shown in Fig. 3(a). When an occluding hair of a skin lesion is misinterpreted as a long and narrow indentation by a pre-processing segmentation program, the lesion border has a high SR, which can be similar to the SR of a much larger indentation as shown in Fig. 3(b). Hence, using the SR measure requires all hairs to be removed carefully either by shaving, or by using a pre-processing program such as DullRazor (Lee et al., 1997).
In this paper, we extend the SR measure to two new area-based measurements, called the most significant irregularity index (MSII) and the overall irregularity index (OII), by directly locating and measuring indentations and protrusions along the lesion border so that the resulting measure is sensitive to structure irregularities. To examine the effectiveness of these new measurements, we performed a user study with fourteen experienced dermatologists by comparing their clinical evaluations of forty lesion borders against the MSII, OII, CI, FD and SFD.
In Section 2, we describe the algorithm for our new measures in detail. Section 3 describes our user study with expert dermatologists. Section 4 discusses the results and Section 5 concludes with a summary and further work.
Section snippets
Preprocessing lesion images
The original skin lesion images are automatically pre-processed by two programs to extract the lesion border contour. First, the skin image is screened for dark thick hairs. These hairs are removed by a software program called DullRazor (Lee et al., 1997) to reduce hair interference on the subsequent automated segmentation program. Then the lesion border contour is extracted automatically (Lee et al., 1995). Fig. 2 shows three of the extracted lesion borders.
Abstraction of lesion border
In order to analyse indentations and
User study
We designed a user study to test the new measures, the MSII and OII, with other common shape descriptors, namely the CI, FD and SFD, using 40 lesion borders which were selected from our pigmented lesion image database. These tested measurements were compared statistically with 14 dermatologists’ clinical evaluations.
Results
The clinical evaluation results of 14 dermatologists are reported in columns ‘C1’ to ‘C14’ in Table 1. The coefficient of concordance, Kendall W, for all 14 clinical evaluations was then determined as W=0.77 (p=0.00). The clinical evaluations were averaged for each lesion border and listed in column ‘Avg’ of Table 2. The tested measurements (CI, FD, SFD, OII and MSII) were also computed and presented in Table 2. The most significant indentation/protrusion segment for each tested border is
Conclusion and future work
We have designed and implemented a new measure called the irregularity index (II) for estimating the border irregularity of melanocytic lesions. The advantage of the new measure is that it directly locates indentations and protrusions along the lesion border. Extending scale-space images to analyse curve segments, we enumerate all global irregular segments and compute their associated indices. This set of measurements provides a rich description for the lesion border that is sensitive to
Acknowledgements
The authors thank Dr. Chantal Bolduc, Dr. Richard Crawford, Dr. William Danby, Dr. Anne Davis, Dr. Brian Gregory, Dr. Vincent Ho, Dr. Harvey Liu, Dr. Lynette Margesson, Dr. Francisco Paschoal, Dr. Jason Rivers, Dr. Jerry Shapiro, Dr. Cecil Sigal, and Dr. David Zloty for their assistance in evaluation the lesion borders and Dr. Andy Coldman and Dr. John Spinelli for their advice on the statistical analysis of the experiment data. Furthermore, this work was supported in part by a BC Health
References (54)
- et al.
A possible new tool for clinical diagnosis of melanoma: the computer
Clinical Diagnosis of Melanoma
(1987) - et al.
Shape analysis for classification of malignant melanoma
J. Biomed. Eng.
(1992) - et al.
Automatic detection of irregular borders in melanoma and other skin tumors
Computerized Medical Imaging and Graphics
(1992) - et al.
DullRazor: A software approach to hair removal from images
Comput. Biol. and Med.
(1997) - et al.
Condon constraints on closed 2D shapes
Computer Vision, Graphics, and Image Processing
(1985) - et al.
Editorial: digital imaging in dermatology
Computerized Medical Imaging and Graphics
(1992) - et al.
Automated feature detection in digital images of skin
Computer Methods and Programs in Biomedicine
(1991) - et al.
Cutaneous malignant melanoma
- et al.
The curvature primal sketch
IEEE Trans. PAMI
(1983) Some informational aspects of visual perception
Psychol. Rev.
(1954)