基于交互引导网络的小样本肠道息肉分割

Interactive Guidance Network for Few-shot Intestinal Polyp

摘要:

肠道息肉分割能够精准定位息肉在结肠中的位置，为临床医生判断是否癌变以及癌变的可能性提供了依据。然而，现有的肠道息肉分割方法仍然存在两个方面的挑战：（1）分割模型的性能过度依赖逐像素标注的训练样本；（2）现有方法对有限标注的样本集利用不充分，往往引入冗余和无关干扰信息，难以应对新出现、新病变等复杂多变的肠道息肉分割问题。为此，利用支持分支和查询分支的双分支网络，设计了一种基于交互引导网络的小样本肠道息肉分割新方法。首先，利用支持图片的真实掩码将支持图片的编码特征划分为目标前景和背景，并在前景特征上计算自注意力，增强目标区域特征的鲁棒性。其次，建立支持前景和查询混合特征的交叉注意力，促进查询混合特征的前景和背景分离，降低查询混合特征中无关背景噪声的干扰。最后，以交叉融合特征生成指导查询分支中未知区域掩码预测的原型集。在bkai-igh-neopolyp、CVC-ClinicDB和EndoTect_2020_Segmentation_Test_Dataset数据集上进行了测试，实验结果验证了所设计方法的优越性。

Abstract:

ntestinal polyp segmentation has been shown to accurately locate the position of polyps in the colon, providing a basis for clinicians to determine whether cancerous changes have occurred and the likelihood of such changes. However, two challenges are still faced by existing intestinal polyp segmentation methods: (1) The performance of segmentation models is overly dependent on pixel-by-pixel annotated training samples; (2) Existing methods do not fully utilize limited annotated sample sets, often introducing redundant and irrelevant interfering information, making it difficult to address complex and variable intestinal polyp segmentation problems such as newly emerging and new lesions. To address these challenges, a new few-shot intestinal polyp segmentation method based on an interactive guidance network has been designed, utilizing a dual-branch network with support and query branches. First, the encoded features of the support image are divided into target foreground and background using the true mask of the support image, and self-attention is calculated on the foreground features to enhance the robustness of the target area features. Second, cross-attention between the support foreground and query mixed features is established to promote the separation of foreground and background in the query mixed features, reducing the interference of irrelevant background noise in the query mixed features. Finally, a prototype set is generated from the cross-fused features to guide the prediction of unknown area masks in the query branch. The superiority of the designed method has been verified through experiments conducted on bkai-igh-neopolyp, CVC-ClinicDB, and EndoTect_2020_Segmentation_Test_Dataset datasets.