SG-UNet：基于全局注意力和自校准卷积增强的黑色素瘤分割模型

打开文本图片集

Abstract：ObjectiveWeproposeanewmelanomasegmentationmodelSG-UNet，toehancetheprecisionofelnoma segmentationindermascopyimagestofacilitateearlymelanomadetection.MethodsWeutlizedaU-shapedconvolutional neuralnetworkUetnddeovttcboneieiosdoslnolngiote backbone，withreferencetothestructureofVGG，weincreasedthenumberofonvolutionsfrom10to13inthedowsampling partofUNettoachieveadeepenednetworkhierarchythatallwedcaptureofmorerefinedfeaturerepresentations.Tofurther enhance featureextraction and detailrecognition，wereplacedthe traditional convolution thebackbonesection with selfcalibratedconvolution toenhancethemodel'sabilitytocaptureboth spatialand channeldimensionalfeatures.Inthepooling part，the original pooing layer was replaced by Haar wavelet downsampling to achieve more effective multi-scale feature fusionandreducethespatialresolutionofthefeaturemap.Theglobalatention mechanismwasthen incorporated intothe skipconnectionsateachlayertoenhancetheunderstandingofcontextualinformationoftheimage.ResultsTheexperimental resultsshowedthattheSG-UNetmodelachieved significantlyimprovedsegmentationaccuracyonISIC2017and ISIC2018 datasetsascompared withothercurrentstate-of-the-artsegmentationmodels，with Dice reached 92.41% and 86.62% and IoU reaching 92.31% and 86.48% onthetwo datasets，respectively.Conclusion The proposed model iscapable of effective and accurate segmentationofmelanoma fromdermoscopy images.

Keywords：imagesegmentation;globalattentionmechanism;melanoma;UNet;self-calibratedconvolution;Haarwavelet downsampling; SG-UNet

黑色素瘤是一种危险的，高度侵袭性的皮肤肿瘤，其特征为快速生长、易于转移和预后不良。（剩余16070字）