基于自适应边界增强和上下文聚合的结直肠息肉分割

Colorectal polyp segmentation based on adaptive boundary enhancement and context aggregation

摘要:

针对结直肠息肉图像中病灶区域边界模糊、尺度变化显著、对比度低等问题，本文提出了一种基于自适应边界增强和上下文聚合的结直肠息肉分割网络模型。首先，以金字塔视觉Transformer作为编码器逐层提取息肉图像中多尺度的远程依赖特征；其次，提出局部全局上下文聚合模块，挖掘局部和全局的上下文语义线索，有效抑制背景噪声干扰；最后，构建渐进式双解码器，通过跨层特征融合模块和两个级联的自适应边界增强模块有效地融合跨层的语义信息，逐级强化边界特征，精细刻画息肉边缘。实验结果显示，模型在CVC-ClinicDB、Kvasir-SEG、CVC-ColonDB和ETIS-LaribPolypDB数据集上的平均Dice系数分别为0.944、0.934、0.817和0.800，平均交并比分别为0.897、0.885、0.736和0.725，分割性能优于现有主流方法，验证了模型的有效性。

Abstract:

In order to solve the problems of blurred boundary, significant scale change, and low contrast of the lesion area in colorectal polyp images, this paper proposes a colorectal polyp segmentation network model based on adaptive boundary enhancement and context aggregation. First, the pyramid visual Transformer is used as the encoder to extract multi-scale long-range dependence features in polyp images layer by layer; secondly, a local and global context aggregation module is proposed to mine local and global contextual semantic clues an-d effectively suppress background noise interference; finally, a progressive The dual decoder effectively fuses cross-layer semantic information through a cross-layer feature fusion module and two cascaded adaptive boundary enhancement modules, gradually strengthening boundary features and finely depicting polyp edges. Experimental results show that the average Dice coefficients of the model on the CVC-ClinicDB, Kvasir-SEG, CVC-ColonDB and ETIS-LaribPolypDB data sets are 0.944, 0.934, 0.817 and 0.800 respectively, and the average intersection and union ratios are 0.897, 0.885, 0.736 and 0.736 respectively. 0.725, the segmentation performance is better than the existing mainstream methods, verifying the effectiveness of the model.