Study On The Algorithms Of Lesion Border Detection In Dermoscopy Images

Melanoma has the highest rate of death in skin cancer,but it is difficult to find in the early times and lose the best treatment time because of the lack of medical diagnosis.Dermoscopy image can effectively help the diagnosis of melanoma,and the border detection of dermoscopy imageis is an important step in computer automatic diagnosis.The existing border detection algorithm can not effectively segment the lesion area,which is characterized by complex texture and irregular shape,and greatly affects the use of computer automatic diagnosis in clinical practice.In order to improve the accuracy and reliability of the border detection algorithm of dermoscopy image,this thesis presents a framework of dermoscopy image border detection algorithm based on superpixel and machine learning,including dermoscopy image preprocessing and super pixel segmentation,superpixel classification by machine learning,post-processing and other steps.The main contents of this thesis include:1.Proposes a set of preprocessing algorithms are researched and implemented,including the removal of image noise and image enhancement.The noise removal algorithm includes the removal of the black frame noise and the removal of the hair noise.The black frame noise and the hair noise are effectively removed through these algorithms,which laid a solid foundation for the extraction and classification.2.Studies and implements the algorithm of lesion border detection based on superpixl.These pixel blocks preserve the effective features identified by the subsequent image boundary detection and do not destroy the boundary information of the lesion area in the image and can effectively improve the accuracy of border detection algorithm.3.Studies and implements feature extraction and classification of superpixel.This paper analyzes the characteristics of dermoscopic images and extracts the four characteristics of superpixel concludes texture,color,gray color difference with background and label distribution of surrounding superpixel,and then classifies superpixels with SVM,and obtains the initial segmentation result of the dermoscopy image.4.Studies and implements classification of superpixels by use of convolution neural network.This paper uses the convolution neural network to automatically extract feature and classify the superpixels of the dermoscopy image.This paper studies theinfluence of different network structure and parameters on the accuracy of the final border detection algorithm.5.Studies and implements a set of post-processing algorithms,including sub-regional merging,islanding,hole filling and border smoothing.Through these algorithms,the image can be divided into two parts: the lesion area and the skin area of the background,which effectively reduces the difference with standard.