Research On Skin Melanoma Image Segmentation Algorithm Based On Deep Learning

Melanoma is a common malignant tumor that mostly occurs in the skin,and its incidence rate accounts for the third place among skin malignant tumors.Due to the nonuniform size,shape,and color of melanoma,it is difficult to segment its border and lesion area.At present,the use of deep learning methods to achieve medical image segmentation has greater advantages in real-time and accuracy than traditional segmentation algorithms.Therefore,this paper proposes to use deep learning method to segment skin melanoma images,the main work is as follows:(1)Aiming at the problem of less image data sets and the image acquisition is easily affected by light,a preprocessing method for medical image data and a postprocessing method for segmentation results are proposed.Among them,preprocessing refers to reducing the influence of light in the data collection process through the color constant algorithm;then amplifying the data can effectively solve the problem of less medical image data.A single image can be enhanced through different data enhancement methods.Increased to 26;the post-processing methods for the segmentation results include conditional random field and test data enhancement,these two methods can refine the segmentation results to varying degrees.Experiments show that preprocessing and postprocessing methods can further improve the accuracy of segmentation.(2)Aiming at the problems that image samples are disturbed by dark skin areas,hair,etc.during the collection process,resulting in large noise,and the small proportion of some melanomas in lesion images,this paper proposes an improved DC-UNet network segmentation algorithm.The algorithm first introduces a channel attention ECA-Net module,so that the model can focus more on the lesion area of melanoma;and then improves the Res-Path module in the network,so that it can better connect the encoder and decoder to Obtain additional image feature details;at the same time,due to the small proportion of melanoma lesion areas in most sample images,in order to improve the accuracy of melanoma boundary segmentation,a weighted cross-entropy loss function is used to overcome the number of classification pixels in the image.The problem of balance,and adding L2 regularization to the loss function to enhance the generalization ability of the model.Compared with the DC-UNet algorithm on the ISIC2017,ISIC2018,PH2 and DSB datasets,the algorithm in this paper improves the segmentation accuracy from 0.9513,0.9444,0.9361,and 0.9781 to 0.9623,0.9537,0.9650,and 0.9847,respectively.Experiments show that the algorithm in this paper has a better segmentation effect than the DC-UNet algorithm,can assist doctors to improve the efficiency of diagnosing melanoma diseases,and can be well applied to other medical image segmentation tasks..(3)Aiming at the problem that skin melanoma has higher requirements for boundary segmentation and some samples in the dataset involve small target segmentation,an improved DeepLab V3+ network segmentation algorithm is proposed.First,the CBAM module is introduced into the Xception model of the backbone network of DeepLab V3+,aiming to allow it to obtain more boundary information to improve the edge segmentation accuracy of the image;secondly,coordinate attention is introduced after the ASPP module in the DeepLab V3+ network to further improve the network The feature extraction ability;secondly,in order to avoid the loss of some semantic information in the process of image feature extraction,the two 4 times upsampling in the network model upsampling process are replaced by two consecutive 2 times upsampling;finally,aiming at the problem that the proportion of lesion area in some images of skin melanoma is small,it is proposed to use Focal Tversky Loss function as the loss function of the algorithm to improve the accuracy of segmentation.The improved network has a segmentation accuracy of 0.9601 and 0.9515 on the ISIC2017 and ISIC2018 datasets,respectively.Compared with the original Deeplab V3+,the improved algorithm has a higher segmentation accuracy of the lesion area,which greatly reduces the segmentation There may be omissions and false detections.