Research On Malignant Melanoma Image Segmentation Method Based On Neutrosophic Theory

Malignant melanoma,as a representative of skin cancer,has a high incidence and mortality rate,and has been growing widely around the world.Currently,skin disease diagnosis and screening largely rely on rich clinical experience of experts,which not only consumes a lot of medical resources but also makes the detection results susceptible to the observer’s visual interpretation and subjective judgment.The aforementioned reasons have made the research on skin lesion segmentation methods for dermoscopy images become an important topic.However,dermoscopy images usually contain complex visual features and a large amount of artifacts.There are two urgent challenges to the research of skin lesion segmentation in dermoscopy images.On one hand,malignant melanoma has an unpredictable and irregular shape,which leads to large intra-class variability and confusion between classes in dermoscopy images,which increases the complexity for model-learning.On the other hand,there is a large amount of uncertainty information in dermoscopy images,making it difficult to obtain accurate neighborhood information of pixels for capturing the lesion boundaries.Therefore,the key problems in malignant melanoma skin lesion segmentation is how to effectively capture the visual features of skin diseases from dermoscopy images and obtain clear pixel neighborhood information.This dissertation aims to identify and extract skin lesion regions from dermoscopy images,focusing on two key aspects:establishing effective mechanisms for extracting dermatological features and obtaining clear pixel neighborhood information.Specifically,the article addresses key technologies for assisting the diagnosis of malignant melanoma:a two-stage classification method combining blue-white veil feature and deep features,shape feature extraction method of malignant melanoma dermoscopy image,single-valued neutrosophic entropy measurement method based on isoentropic cylindrical surface,and neutrosophic representation method for malignant melanoma image segmentation.The experimental analysis shows that these methods can effectively improve the accuracy of skin lesion classification and segmentation.The main contributions and innovations of this dissertation are as follows:(1)Aiming at the insufficient analysis of blue-white veil information of malignant melanoma,a classification method based on blue-white veil feature and depth features of dermoscopic images is proposed.Specifically,this method utilizes the extension correlation function to quantify the difference in the distribution range of different objects,establishes the extension distance and extension location value under the two-interval nesting,divides the classical field and controlled field of pixels with different image structures,and finds the most important color feature of malignant melanoma,namely the blue-white veil,for the first stage classification of skin lesion images.The YoDyCK model is also proposed to quickly extract the potential features of the lesions gathered in the center of the dermoscopy images,then these deeper and more abstract features are used for the second stage classification of skin lesion images.Experimental results show that the proposed method can effectively extract significant features of malignant melanoma from dermoscopy images,thereby improving the classification accuracy of malignant melanoma dermoscopy images.(2)To address the problem that the visual features in the medical field are difficult to be effectively represented and quantified,a shape feature extraction method for malignant melanoma based on neutrosophic segmentation is proposed to improve the accuracy of malignant melanoma diagnosis.Specifically,this method establishes a fuzzy representation of skin lesion images by using the neutrosophic theory to enhance the differences between the lesion edge and other pixel structures,and capture the lesion shape of skin diseases.Based on this,the method maps the lesion shape into a saliency curve and a series of inner diameters of the skin lesion,which reduces the limitations of analyzing lesion boundary information and quantifies the symmetry and uniformity of the skin lesion boundary.According to the dermatological similarities and differences between black nevi and malignant melanoma,this method uses the edge of nevi as a reference to measure the regularity of the contour of skin lesions.Combining the blue-white veil information of the lesion extracted with extension theory,a set of visual features of skin lesion images is established.Then the feature set is inputted into a support vector machine to optimize the results of dermoscopy images in first-stage classification.Experimental results demonstrate that the malignant melanoma features extracted by this method can effectively improve the classification accuracy of malignant melanoma dermoscopy images.(3)To address the problem that the spatial neighborhood information of pixels in dermoscopy image is difficult to obtain effectively,an measurement method for skin lesion pixel information uncertainty based on neutrosophic entropy is proposed to highlight the edge of the lesion with the most uncertain pixel information in the dermoscopy image.Specifically,this method combines the use of the neutrosophic theory to establish a fuzzy representation of skin lesion images.By using three independent components of true membership,uncertain membership and false membership,this method effectively represents the degree of certainty and uncertainty of pixels.Based on this,the method constructs a novel neutrosophic entropy measurement model,namely,the neutrosophic entropy measurement method based on isoentropic cylindrical surface.This method projects image pixels to the isoentropic cylindrical surface to establish an effective spatial relationship between pixels,capturing the most uncertain lesion edge in dermoscopy images.Experimental results show that this method effectively improves the difference between the lesion edge and the healthy skin and the pixels inside the lesion.(4)Aiming at the problem that the multi-color non-uniform mixed distribution of pixels in malignant melanoma dermoscopy images leads to fuzzy structural boundaries and uncertain pixel region attribution,a neutrosophic representation method for malignant melanoma image segmentation is proposed in this dissertation to optimize the neutrosophic conversion process of the image and further highlight the irregular edges of the lesion.Specifically,this method separates the color and texture information of each channel in the image,and obtains the determined shape and uncertain boundary of the image pixel structure.Based on this,the true membership degree and uncertain membership degree of pixels are established,and the neutrosophic images in R,G and B channels are established to observe the shape of lesion area from different angles.Moreover,this method establishes the optimal channel combination image of the neutrosophic image,and the image information is fully utilized to further increase the difference between the edge of the lesion and other areas,and the reliability of the segmentation method is improved.Meanwhile,the target edge in the image is smoothed by morphological method,and the image is segmented by clustering method to assist doctors in determining the excision boundary of the lesion.Experimental results show that this method can effectively extract the banded edges of malignant melanoma,and improve the segmentation accuracy of malignant melanoma dermoscopy images compared with other methods.