融合局部和全局特征的息肉分割模型

Polyp segmentation model based on fusion of local and global features

针对现有模型在息肉分割中存在复杂区域分割困难、边缘细节信息丢失、泛化能力不足等问题,提出一种融合局部和全局特征的息肉分割模型。以卷积神经网络和Transformer作为并行编码器,使模型可以兼顾多种尺度的局部细节特征和全局语义特征;在跳跃连接处构建注意力增强模块和多尺度残差模块,前者强化模型对重要信息的关注度,后者高效探索目标区域并准确预测其边界,同时促进不同层次特征之间的交互;在解码阶段采用基于残差的逐步上采样特征融合方式汇聚各阶段特征,进一步增强模型的感知能力,丰富息肉特征;最后使用高效预测头促进浅层特征的融合,输出分割结果。该模型在多个对比实验中表现最优,同次优模型相比,在Kvasir、CVC-ClinicDB数据集上,mDice平均提升了1.21%;mIoU平均提升了1.82%;在CVC-ColonDB、ETIS数据集上,mDice平均提升了2.67%,mIoU平均提升了2.83%。实验结果表明,相比于现有主流模型,该模型具有较优的分割精度和泛化性能。

To solve the problems of difficulty in segmentation of complex areas,loss of edge details,and insufficient generalization ability in polyp segmentation by existing models.This paper proposed a polyp segmentation model based on fusion of local and global features.Convolutional neural network and Transformer are used as parallel encoders to make the model take into account both the local detail features and global semantic features of multiple scales.The attention enhancement block and the multi-scale residual block are constructed at the jump junction.The former enhances the model′s focus on important information,while the latter efficiently explores the target regions and accurately predicts theirs boundaries,while promoting the interaction between different levels of features.The residual-based stepwise upsampling feature fusion method is used in the decoding stage to gather the features of each stage,which further enhanced the perception ability of the model and enriched the polyp features.Finally,the efficient prediction head is used to promote the fusion of shallow features and output the segmentation results.The model performs best in the comparative experiments.Compared with the sub-optimal model,on the Kvasir and CVC-ClinicDB datasets,it achieved an average mDice improvement of 1.21%and an average mIoU improvement of 1.82%;on the CVC-ColonDB and ETIS datasets,it achieved an average mDice improvement of 2.67%and an average mIoU improvement of 2.83%.The experimental results show that the proposed model has better segmentation accuracy and generalization performance than the existing mainstream models.