Biologically Inspired Quad Tree Color Detection in Dermoscopy Images of Melanoma
Abstract: This paper presents a Quad Tree-based melanoma detection system inspired by dermatologists’ color perception. Clinical cooler assessment in dermoscopy images is challenging because of subtle differences in shades, location-dependent color information, poor color contrast and wide variation among images of the same class. To overcome these challenges, color enhancement and automatic color identification techniques, based on Quad Tree segmentation and modeled after expert color assessments, are developed. The approach presented in this paper is shown to provide an accurate model of expert color assessment. Specifically, the proposed model is shown to: (i) identify significantly more colors in melanomas than in benign skin lesions, (ii) identify a higher frequency in melanomas of three colors: blue-grey, black and pink, and (iii) delineate locations of melanoma colors by quintiles, specifically predilection for blue-grey and pink in the periphery and a trend for white and black in the lesion centre. Performance of the proposed method is evaluated using four classifiers. The kernel support vector machine (SVM) classifier is found to achieve the best results, with an area under the receiver operating characteristic (ROC) curve of 0.93, compared to average area under the
ROC curve of 0.82 achieved by the dermatologists in this study. The results indicate that the biologically inspired method of automatic color detection proposed in this paper has the potential to play an important role in melanoma diagnosis in the clinic. Existing system: Color is a critical feature in melanoma, especially when combined with location in the lesion in capturing color chaos and variegation. Analytic color descriptors usually employ the red, green and blue (RGB) color space. These descriptors are among the most significant of all analytic descriptors of melanoma. Ferris et al. found that the top three statistical features for melanoma were all colour descriptors from colour histograms and colour asymmetry. Rubegni et al. found that the most significant feature for melanoma detection was red asymmetry, capturing the “colour island� chaos. Other colour features for melanoma detection are variation of hues , analytical colour analysis of variegation , RGB color channel statistical parameters , spherical and La*b* color coordinate features , and the number of colours of concern present within the skin lesion . Colour quantization was performed using the median split colour algorithm. Crisp and fuzzy color histogram techniques were used to identify melanoma colours. Proposed system: Adding the input dermoscopy image to the top-hat filtered image to enhance bright features, then subtracting the bottom-hat filtered image from the input image to enhance dark features, using an optimized structuring element to probe the input image. The input image is a (L channel) grayscale and the structuring element (b) is a small matrix of pixels, each with a value of zero or one that can have any arbitrary shape and size. In the proposed method, we use a disk-shaped b with radius of 18 to apply morphology transforms, for more details see. Hair is removed with directional morphological operations, filtering, and interpolation Images are smoothed by a Gaussian low-pass filter. Advantages: Adding the input dermoscopy image to the top-hat filtered image to enhance bright features, then subtracting the bottom-hat filtered image from the input image to
enhance dark features, using an optimized structuring element to probe the input image. The input image is a (L channel) grayscale and the structuring element (b) is a small matrix of pixels, each with a value of zero or one that can have any arbitrary shape and size. Discriminatory features developed by Dalal et al. Often used a pair of adjacent deciles, supporting the idea that deciles are too narrow. To construct the first quintile overlay, the distance transform is applied successively to pixels, beginning with those adjacent to the background; Disadvantages: Melanoma classification is a binary decision problem. All 320 extracted features described above are operated upon by 4 different classifiers (Kernel support vector machines (SVM), backpropagation neural network (ANN), linear discriminant analysis (LDA) and random forests) using the corresponding parameter ranges shown in Table 1 to ascertain whether the lesion is a melanoma or benign. Colour contrast enhancement increases the primary contrast between lesion and non-lesion areas, but can decrease colour contrast within the lesion. To remedy this problem, the top-hat and bottom-hat morphological transformations are combined to correct the effects of non-uniform illumination. Modules: Colour analysis: Colour is a critical feature in melanoma, especially when combined with location in the lesion in capturing colour chaos and variegation. Analytic colour descriptors usually employ the red, green and blue (RGB) colour space. These descriptors are among the most significant of all analytic descriptors of melanoma. Ferris et al. found that the top three statistical features for melanoma were all colour descriptors from colour histograms and colour asymmetry. Rubegni et al. found that the most significant feature for melanoma detection was red asymmetry, capturing the “colour island� chaos. Other colour features for melanoma detection are variation of hues, analytical colour analysis of variegation, RGB colour channel statistical parameters, spherical and La*b* colour coordinate features, and the number of colours of concern present within the skin lesion. Colour quantization
was performed using the median split colour algorithm. Crisp and fuzzy colour histogram techniques were used to identify melanoma colours .In this paper, Quad Tree colour segmentation is used to identify critical colours modeled by dermatologists. A classifier is employed to use colour shade, texture and location information; and its performance is compared to that of dermatologists. Constructing Quintile Overlays : Dermatologists use a two-phase model of melanoma growth, with the earliest radial growth phase (expanding outward) followed by a vertical growth phase (expanding downward). This has motivated researchers to analyze melanomas by concentric quartiles, or deciles, using the Euclidean distance transform to construct concentric overlays . Dermatologists in the clinical phase of this paper suggested quintiles to be a better location descriptor, to avoid too-narrow deciles and toowide quartile descriptors. Discriminatory features developed by Dalal et al. Often used a pair of adjacent deciles, supporting the idea that deciles are too narrow. To construct the first quintile overlay, the distance transform is applied successively to pixels, beginning with those adjacent to the background; when the transform has included one-fifth of lesion pixels, the outer quintile is determined; successive quintiles are determined similarly. A dermoscopy image with quintile overlays. Quad Tree Clustering in Palette: For colour recognition, each quintile lesion is divided into MĂ—M square blocks. A block is characterized by having similar colour features, as described previously. In order to obtain a single representative colour, it is necessary to use a meaningful measure of similarity within the colour patches. Since prior work determined that mean colours in the CIELAB colour space had optimal discriminative power across all colour patches , the image is converted to CIELAB color space ;all patches are described in CIELAB colour space. Using mean and entropy values, each block is merged to the nearest neighbor, using the Euclidean norm; each block of the image is mapped into one of the six colours comprising the colour palette. After assigning all blocks to their corresponding patches, those belonging to the same colour group are merged to form the region corresponding to that particular colour (Fig. 3d and Fig.4 (centre)) and then the final image converted
back to RGB format. The final step in colour labeling is to decide whether a colour is present in a clinically relevant amount in a lesion. Classifiers Chosen: Melanoma classification is a binary decision problem. All 320 extracted features described above are operated upon by 4 different classifiers (Kernel support vector machines (SVM), backpropagation neural network (ANN), linear discriminant analysis (LDA) and random forests) using the corresponding parameter ranges shown in Table 1 to ascertain whether the lesion is a melanoma or benign. The choice of classifiers was based on simplicity and performance when used in computer-based melanoma recognition systems. In the following section, we report the results obtained from four distinct experiments.