Research On Skin Lesions Detection Method Based On Dermoscopy Images

Malignant melanoma (MM) is one of the rare skin cancers with an increasing incidence rate. In the U.S. alone, the number of new cases and deaths associated with MM in 2010 are estimated to be 68,130 and 8,700, respectively. In recent years digital dermoscopy has revealed an innovative dimension of clinical morphology in pigmented skin lesions, becoming one of the most cost-effective and non-invasive techniques for early detection of skin cancer. The diagnosis of skin lesions is frequently performed according to the ABCD (A:Asymmetry, B:Border, C:Color, D:Differential structures), Menzies's method,7-point checklist and patterns classification. In particular, it is very difficult to classify the lesions and even experienced dermatologists have diagnostic accuracy below 85%.To improve the clinical performance of dermatologists, automated computer-aided detection (CAD) of lesions tools are designed to provide "second opinion" for improving accuracy, efficiency and consistency of detecting and classifying lesions. Since CAD tools have usually five stages:(1) preprocessing, (2) artifact removal, (3) segmentation of lesions, (4) quantification of ABCD and extract texture-related features for pattern classification, and (5) finally classification. However, the effectiveness of current tools is limited in order to detect and classify skin lesions due to a variety of lesion shapes, irregular boundaries, specular reflection, low-contrast, artifacts, and non-effective pattern classification methods. Therefore, detection of lesions remains a challenging task.The detection of skin lesions, as well as to increase the detection accuracy, enhancement, unsupervised skin lesion segmentation, numerical quantification of lesions, and an effective pattern classification method are studied in detailed in this dissertation. For lesion detection, a new preprocessing step is proposed to perform the adjustment of specular reflection and the contrasts enhancement in uniform color space based on homomorphic transforms filter (HTF) and contrast adjustment methods, respectively. The experimental results indicate that this technique is outperformed than current enhancement methods without degrading the image quality. To repair artifacts such as dermoscopy-gel, a recursive median filtering was used. Next to detect and repair the hair-like artifacts, a new idea about hair-like repaired algorithm is proposed by using 2D derivative of Gaussian (DOG), morphological functions and fast marching inpainting methods. This hair-like artifact repaired algorithm has superior results than state-of-the-art techniques without disturbing the patterns of lesions. Afterwards, a novel idea about the segmentation of lesions is presented based on an improved region-based active contour (IRAC) model. The IRAC model has controlled many disadvantages in the current active contour based segmentation algorithms such as:level set initialization, fixed regularization parameters, and overlapping of the contours in the presence of multiple objects. The experimental results suggest that the IRAC model is outperformed than the state-of-the-art segmentation methods. After detecting the lesions'border, the extraction of a set of quantification features is performed for effectively utilizing the ABCD rule. Moreover, a novel idea about pattern classification is developed. To address the problem of multicomponent patterns that many researchers have forgotten to consider it in their classification systems, an adaptive boosting multi-label learning algorithm (AdaBoost.MC) is developed based on decision trees, maximum a posteriori (MAP) and the robust ranking principles. The proposed method for multi-label learning algorithm is obtained higher accuracy when compared to current learning algorithms without generating the classes'correlation problem.The researches on automatic detection of skin lesions based on preprocessing, artifact removal, segmentation of lesion, lesion quantification and pattern classification techniques are developed in a perceptually uniform color space for providing help to dermatologists towards an increase the sensitivity and specificity for diagnosing the lesions in clinical practice, which could be used to increase the performance level of CAD or content-based image retrieval (CBIR) systems.