Dermoscopic Image Segmentation Algorithm Based on Multi-scale Information Extraction and Feature Fusion
TANG Jianan
MENG Xiangrui
Abstract：Aiming at the low segmentation accuracy of existing dermoscopic image segmentation technologies,we proposed a U-type network model based on multi-scale information extraction and feature fusion U-shaped network(MF-UNet).On the basis of U-Net,a batch normalization layer was added after the convolution layer,and the original skip connection part was replaced by a four-level feature fusion block to make full use of semantic information and location information.A multi-scale atrous convolution block and a multi-scale pooling block were added at the end of the feature extraction end to increase the receptive field,and a two-path concating upsampling block was used for upsampling which reduced the information loss in the process of image recovery.The experiments showed that compared with the U-Net model,the mean intersection over union(MIoU)of MF-UNet increased by 14.32%,and Dice similarity coefficient(DSC)increased by 13.18%.Good results were obtained.This study provided some reference for the work of computer technology assisting doctors in the diagnosis of dermatosis.
Keywords：semantic segmentationdermoscopic imagefeature fusionattention mechanismMF-UNetdeep learninglesion margin
Publication Date：2024-06-20
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：7( 226-232 )