梁礼明, 何安军. 基于Swin Transformer和图形推理的结直肠息肉分割方法[J]. 工程科学学报. DOI: 10.13374/j.issn2095-9389.2023.04.21.004
引用本文: 梁礼明, 何安军. 基于Swin Transformer和图形推理的结直肠息肉分割方法[J]. 工程科学学报. DOI: 10.13374/j.issn2095-9389.2023.04.21.004
Colorectal polyp segmentation method based on Swin Transformer and graph reasoning[J]. Chinese Journal of Engineering. DOI: 10.13374/j.issn2095-9389.2023.04.21.004
Citation: Colorectal polyp segmentation method based on Swin Transformer and graph reasoning[J]. Chinese Journal of Engineering. DOI: 10.13374/j.issn2095-9389.2023.04.21.004

基于Swin Transformer和图形推理的结直肠息肉分割方法

Colorectal polyp segmentation method based on Swin Transformer and graph reasoning

  • 摘要: 针对结直肠息肉图像分割中病灶区域尺度变化大、边缘模糊以及息肉与正常组织对比度低等问题,导致病变区域分割精度低和分割边界存在伪影,提出一种基于Swin Transformer和图形推理的自适应网络。该网络一是利用Swin Transformer编码器逐层提取输入图像的全局上下文信息,多尺度分析病变区域的显著性特点。二是提出全局与局部特征交互模块增强网络对复杂病灶的空间感知能力,突出待分割目标的关键位置信息。三是通过区域引导图推理模块以图循环递推的方式挖掘先验信息之间的高阶显性关系。四是设计面向边缘细节的边缘约束图推理模块,整合边缘细节,改善分割效果。在CVC-ClinicDB、Kvasir、CVC-ColonDB和ETIS数据集上进行实验,其Dice系数分别为0.939,0.926,0.810和0.788,平均交并比分别为0.889,0.879,0.731和0.710,分割性能优于现有方法。仿真实验结果表明,对于形态结构复杂、对比度低和边缘模糊的结直肠息肉图像均有较高的分割精度。

     

    Abstract: Aiming at the problems of large scale changes of the lesion area, blurred edges, and low contrast between polyps and normal tissues in the image segmentation of colorectal polyps, which lead to low segmentation accuracy of lesion areas and artifacts in the segmentation boundary, an automatic segmentation algorithm combining Swin Transformer and graph-line reasoning is proposed. To adapt to the network, the network first uses the Swin Transformer encoder to extract the global context information of the input image layer by layer, and analyzes the salient characteristics of the lesion area at multiple scales. The second is to propose a local global feature interaction module to enhance the network's spatial perception of complex lesions and highlight the key location information of the target to be segmented. The third is to use the region-guided graph reasoning module to mine the high-order explicit relationship between prior information in the way of graph cycle reasoning. The fourth is to design an edge-constrained graph reasoning module oriented to edge details, which integrates edge details and improves the segmentation effect. Experiments were carried out on the CVC-ClinicDB, Kvasir, CVC-ColonDB and ETIS datasets, the Dice coefficients were 0.939, 0.926, 0.810 and 0.788, and the average cross-merge ratios were 0.889, 0.879, 0.731 and 0.710 respectively. There is a way. Simulation results show that the segmentation accuracy is high for colorectal polyp images with complex morphology, low contrast and blurred edges.

     

/

返回文章
返回