摘要：

结肠息肉具有边界不清晰且大小、颜色、形状各异的特点，使得采用深度学习方法提高其分割性能仍是一项极具挑战性的工作。为提高息肉分割的准确率，提出一种基于多任务联合注意力的结肠息肉分割网络CPMJA-Net。为改善Transformer缺乏机制来增强局部区域信息交换的问题，设计级联融合模块以增强网络的局部特征表示，有助于息肉边缘的识别和恢复。受多头Self-Attention机制的启发，构建多任务注意力模块，采用渐进式融合的方式将不同模块得到的特征图逐步进行融合，以凸显关键信息并抑制干扰信息。为更好聚合图像的高级和低级特征，设计联合注意力模块，利用高级特征的轮廓信息在低级特征中筛选出有利于边缘分割的细节特征，并将其与息肉轮廓聚合起来，得到更加精确的边缘分割结果。实验结果表明，CPMJA-Net在4个公开数据集上的性能表现均为最优，与次优的算法相比，CPMJA-Net的mDice系数分别在Kvasir、CVC-CilinicDB、CVC-ColonDB和ETIS数据集上提升0.7、0.8、0.4、0.4个百分点，平均交并比（mIoU）也分别提升1.6、1.2、0.6、0.5个百分点，其能改善过分割问题，弥补注意力机制的不足，提升解码器的细节恢复能力。

关键词: 肠道息肉检测, PVT网络, 自注意力机制, 多任务注意力, 联合注意力

Abstract:

Colon polyps have the characteristics of unclear boundaries and varying sizes, colors, and shapes, making it difficult to improve their segmentation performance using deep learning methods. A colon polyp segmentation network CPMJA-Net based on multi-task joint attention is proposed to improve the accuracy of polyp segmentation. To address the problem of Transformers lacking mechanisms to improve local information exchange, a cascaded fusion module is designed to enhance the local feature representation of the network, which aids in the recognition and restoration of polyp edges.Inspired by the multi-head Self-Attention mechanism, a Multi-task Attention Module(MAM) is constructed, and the feature maps obtained from different modules are gradually fused using a progressive fusion method to highlight key information and suppress interference information. A Joint Attention Module(JAM) is designed to use the contour information of advanced features to filter out detailed features that are conducive to edge segmentation from low-level features and aggregate them with the contour of polyps to obtain more accurate edge segmentation results to better aggregate the advanced and low-level features of images. The experimental results show that CPMJA-Net has the best performance of all four public datasets. Compared with the suboptimal algorithm, the mDice coefficient of CPMJA-Net has improved by 0.7, 0.8, 0.4, and 0.4 percentage points on the Kvasir, CVC-CilinicDB, CVC-ColonDB, and ETIS datasets, respectively. In addition, the mean Intersection over Union(mIoU) increased by 1.6, 1.2, 0.6, and 0.5 percentage points, respectively. Experiments have shown that CPMJA-Net improves over-segmentation, compensates for attention mechanism shortcomings, and improves the decoder's ability to recover details.

Key words: detection of intestinal polyps, PVT network, self-attention mechanism, multi-task attention, joint attention