Research On Skin Melanoma Segmentation Method Based On Convolution Neural Network

Melanoma is a highly malignant tumor formed by abnormal accumulation of melanocytes on the skin,which grows rapidly and is easy to transfer.In recent years,the incidence rate and mortality of cutaneous malignant melanoma have gradually increased,which has become a serious public health problem and aroused widespread concern in society.Artificial intelligence technology based on deep convolutional neural networks has made significant achievements in the application of automated segmentation of skin melanomas.However,due to the diverse manifestations of pathological features of skin lesions and plaques,uneven edges,and most pathological slices belong to ultra-high resolution images,the computational accuracy of automated segmentation algorithms for skin melanomas is not high,and it consumes huge memory of GPU.Based on this,the main work of this article is as follows.(1)A two-stage fine segmentation algorithm GLNF with low memory consumption of GPU is proposed.The global segmentation network adopted in the first stage of the algorithm improves the feature pyramid structure with Res Net50 as the backbone.In the process of image feature extraction,the global pyramid average pooling module is used to enhance the extraction of image global semantic information;The multi-scale feature fusion branch is adopted to integrate the semantic information of the high-level feature map into the lower level feature map to enhance the representation ability of the semantic information of the low-level feature map.In the second stage,a global to local fine segmentation strategy is adopted.The image is clipped based on the global segmentation results to obtain a smaller candidate area,which is input into the local segmentation network.The local segmentation network only processes pixels in the candidate region and shares image features with the corresponding layer of the global network.The memory consumption of GPU is minimized while refining the segmentation results.Validate GLNF on the classic dataset ISIC2018.The experimental results show that the accuracy of the algorithm is the highest compared with other classic segmentation algorithms,and the memory of GPU is significantly reduced,which can be effectively applied to the segmentation task of high-resolution skin lesions image.(2)An improved method PEGLNF based on feature path enhancement is proposed.Although GLNF has achieved a good balance between computational performance and hardware requirements,its computational accuracy can still be further improved.Using a feature path enhancement method,a bottom-up feature path enhancement module is added to the global segmentation network.A shallower convolutional network is used to resample the shallow feature map,preserve as much spatial information as possible in the shallow feature map,and smooth the fused features to increase the accuracy of the model’s segmentation of boundary and shape.The method PEGLNF was also validated on the dataset ISIC2018.The visualization experiment results and various algorithm evaluation indicators prove that PEGLNF has better judgment on the shape and boundary information of the lesions,and the accuracy of the algorithm is higher.