Research On Lesion Segmentation Method Based On Skin Image

The social status quo of insufficient medical resources has urged the emergence of computer-aided diagnosis technology.The computer-aided diagnosis technology of skin diseases has become a major research focus in the combination of medical images and intelligent computing,and the detection and segmentation of lesions is a key step in assisted diagnosis.Skin disease images can be divided into two categories: clinical images and pathological images.This article uses clinical images as the research object,combined with deep learning to study the problem of lesion segmentation.The clinical images of skin diseases are obtained by directly photographing the surface of the lesion with a photographing device,so the image quality obtained is low.For example,most of the images acquired by professional equipment dermoscopes have black frames,shadows and other noise;the images collected by ordinary photographing equipment are mostly affected by light;When the lesion is located in a hairy area,there are natural noises such as hair and blood vessels;In addition,there is a large amount of infiltration and rendering in the skin disease lesions,so there is no clear boundary between the background area of the image and the lesion.Aiming at the problem of lesion segmentation in dermoscopic images,this article will work from the following aspects:Preprocessing the dermoscopic image,including removing black frame noise,removing hair noise,and enhancing the contrast of the foreground and background of the image.The black frame removal method calculates the position of the black frame pixel by pixel and deletes the rows and columns that are the black frame;A threshold-based hair removal method is proposed.The threshold is used to divide the hair into two categories: coarse and sparse.Thick and dense hair is repaired by partial differential operation,and sparse hair is repaired by bilinear interpolation;According to the principle of visual imaging,the contrast enhancement method treats the original image as the product of the incident light and the object itself,and decomposes the object itself by removing the incident light to enhance the edge features of the image and increase the brightness of the image.The neural network is used as the segmentation model to segment the preprocessed image.After studying the application of neural network in lesion segmentation,the U-net model is analyzed,and the U-net model is used to segment skin lesions.In the experiment,the Dice loss function is used as the objective function to improve the U-net model.Experimental results show that the original U-net model can reach 85.21% in the segmentation of skin lesions,while the performance of the model is not greatly improved when Dice is used as the objective function,and the accuracy is only 1.88% higher than the original model.Based on the U-net model,a codec segmentation model based on separable convolution is constructed.The feature extractor of the coding part is composed of separable convolution and hole convolution;The decoding part continues to use the up-sampling operation of the U-net model;The objective function of the segmentation model uses the Dice loss function.The experimental results show that the model can achieve better results in the segmentation of skin lesions,with an accuracy rate of 95.24%,which is 6.17% higher than the U-net model.