首页|基于特征交互融合的结肠息肉图像分割算法研究

基于特征交互融合的结肠息肉图像分割算法研究

扫码查看
针对结肠息肉图像中病灶区域尺度多变,传统方法难以捕捉其内部复杂关系导致的分割精度不高的问题,文章提出一种特征交互融合分割网络FIFNet.在该网络中,利用金字塔式Transformer和ResNet18 并行提取息肉图像的局部与全局特征并通过语义协调单元SH融合两者之间的语义信息;此外,设计了层间注意力聚合模块IA,自适应加权融合不同层级特征,从而突出息肉图像的形态和纹理信息;最后,反向残差注意力模块IRA充分挖掘息肉区域与边界的联系,提高了分割结果的准确性.在公共数据集Kvasir、CVC-ClinicDB、CVC-ColonDB、ETIS、Endosece上进行实验测试,其中mDice系数分别为 0.929、0.941、0.821、0.794、0.900.实验结果表明FIFNet网络在息肉图像分割上具有一定的应用价值.
Research on Colon Polyp Image Segmentation Algorithm Based on Feature Interactive Fusion
A feature interactive fusion segmentation network called FIFNet is proposed in this paper to address the issue of low segmentation accuracy caused by the variable scale of the lesion area in colon polyp images and the difficulty of capturing its internal complex relationships of traditional methods.In the network,it utilizes Pyramid Vision Transformer(PVT)and ResNet18 to extract local and global features of polyp images in parallel,and fuses the semantic information between the two by the Semantic Harmonization(SH)unit.Then,it designs the Inter-layer Attention(IA)aggregation module to highlight the morphological and texture information of polyp images through adaptive weighted fusion of different layer features.Finally,the Inverse Residual Attention(IRA)module fully explores the connection between the polyp area and boundary to improve the accuracy of segmentation results.Experimental tests are conducted on the public datasets Kvasir,CVC-ClinicDB,CVC-ColonDB,ETIS and Endosece where the mDice coefficients are 0.929,0.941,0.821,0.794,and 0.900,respectively.Experimental results show that the FIFNet network has a certain application value in polyp image segmentation.

colon polyp segmentationfeature interactive fusionSemantic Harmonization unitInter-layer AttentionInverse Residual Attention

陆鹏

展开 >

安徽理工大学 计算机科学与工程学院,安徽 淮南 232001

结肠息肉分割 特征交互融合 语义协调单元 层间注意力 反向残差注意力

2024

现代信息科技
广东省电子学会

现代信息科技

ISSN:2096-4706
年,卷(期):2024.8(24)
  • 1