Research On Melanoma Image Segmentation Algorithm Based On Multi-stage Feature Extraction And Lightweight Network

Nowadays,the age of artificial intelligence has arrived.Using computer vision cognitive technology to solve the problem of identifying and labeling lesions in medical images is one of the very successful smart medical application scenarios.As a kind of skin cancer,melanoma has the characteristics of being difficult to detect in the early stage,difficult to treat in the late stage,and having a high mortality rate.As the number of patients gradually increases,it seriously endangers human life,health and safety.Melanoma segmentation often requires doctors to have professional medical knowledge,which is inefficient and costly.Therefore,the use of computer technology to help solve the problem of melanoma segmentation is great significance to improve the diagnosis of skin cancer.In the process of collecting images using dermoscopy,due to the use of oil immersion,incident angle lighting and other technologies,the images appears different brightness,bubbles and other noise.Moreover,the melanoma images has hair covering,fuzzy boundaries,irregular shapes and other reasons,which affect the final segmentation accuracy.Existing work mainly improves the model’s ability to recognize pixels in the lesion boundary area through the following aspects: 1)During the encoding phase,reduce the number of downsampling and increase the resolution of the feature map to reduce missing pixels.However,blindly expanding the resolution of the feature map in the feature extraction phase will bring more redundant information and increase the demand for computing resources exponentially.2)Expand the resolution of high-level feature maps in the decoding phase to restore more details and improve the granularity of the network.But the deconvolution operation will bring grid artifacts.3)Improve the loss function and strengthen the ability to monitor the boundary pixels.In order to improve the accuracy and performance of the melanoma segmentation model,this thesis addresses the above issues:This thesis addresses the problems of inaccurate pixel prediction in the boundary area and the large demand for computing resources in the existing models.Use the dalition residual network and feature pyramid attention mechanism to improve the semantic extraction ability of the network.In the dalition residual network,the dalition convolution is used instead of the max pooling,and the resolution of the output feature image will not be reduced when the receptive field is increased.In addition,for high-resolution dense pixel prediction tasks,in order not to increase the computing resource requirements of the model,feature reuse is used to improve the efficiency of the use of redundant information in the feature map.This thesis uses a progressive up-sampling encoding-decoding structure.The features in the encoding stage repeatedly capture more scale information through the chain structure in the chained residual dilation expansion convolution unit.And use the extracted features to make up for recovering more details during the upsampling process at the decoding stage.The experimental results show that the multi-stage feature extraction melanoma segmentation algorithm designed in this thesis has better pixel discrimination ability than the existing algorithms.Compared with the existing algorithm under the same conditions on the ISIC2017 data set,we achieved the best Dice accuracy of 81.6%.Based on the multi-stage feature extraction network designed in Chapter 3,this thesis designs a low-weight melanoma segmentation network model with less computation and lower memory requirements.According to convolution factorization principle,the following substitutions are made to the algorithm in Chapter 3: 1)Use small convolution kernels instead of large convolution kernels,2)Use depthwise separable convolutions and factorization convolutions instead of ordinary convolutions,3)Use factorization convolutions with holes instead of dalition convolutions,4)Control the number of convolution kernels and reduce the number of parameters according to the channel split-transformation-merge strategy.After the ablation experiment analysis of the model in this thesis,the parameter amount is only 1/5 of the original,and the calculation amount is reduced to 1/47 of the original when the accuracy loss of the algorithm is small.