Abstract: The deep learning method has achieved remarkable success in polyp image segmentation. However, there are still two issues hindering the advancement of high-precision polyp recognition. Firstly, there exists a disparity in the distribution of image features collected from different devices, posing a challenge of domain shift across various datasets. Secondly, the prevailing models predominantly focus on handling features of the same scale, limiting the capacity to capture multi-scale features. To address these issues, a domain adaptation module and a multi-scale feature aggregation module are proposed. The domain adaptation module achieves adaptation to the variances among images derived from distinct domains via an unsupervised approach. Then, the domain-adapted images are processed through an encoder to obtain feature maps with different receptive fields. By employing the proposed multi-scale feature aggregation module, the feature maps with difference receptive fields are fused, enhancing the  ability of model to segment lesions of different scales. On five public polyp segmentation datasets, the proposed method is compared with widely used colonic polyp segmentation methods. The proposed method achieves better results on the Dice and IoU metrics compared to the contrastive methods, particularly on the Kvasir and ClinicDB datasets. Furthermore, the model exhibits consistent segmentation performance employed for assessing generalization ability. In conclusion, the fusion of the domain-adaptive network and the multi-scale feature aggregation in the polyp segmentation network effectively facilitates the segmentation of polyp images while demonstrating commendable generalization performance.

Key words: polyp segmentation, domain adaptation, multi-scale feature aggregation, medical image processing

摘要： 基于深度学习的方法在息肉图像分割上取得了巨大成功，然而仍存在两个问题阻碍高精度息肉识别的发展。第一，不同设备收集的图像在特征分布上存在差异，使得不同息肉分割数据集存在域偏移的问题。第二，现有的模型专注于处理相同尺度大小的特征，限制了模型对多尺度特征的捕捉能力。为解决这些问题，提出了一个域自适应模块和一个多尺度特征聚合模块。域自适应模块采用无监督的方式，自适应不同域图像之间的偏移。将域自适应后的图像输入编码器，获取不同感受野大小的特征图，利用提出的多尺度特征聚合模块，将具有不同感受野的特征图进行聚合，提高模型对不同尺度病灶的分割能力。在五个公开的息肉分割数据集上，与使用广泛的结直肠息肉分割方法进行比较。在Kvasir和ClinicDB数据集上，提出的方法在Dice和IoU指标上与所对比的经典分割方法相比，取得了更好的结果。在验证模型泛化性能的数据集上，依旧表现出稳定的分割性能。综上所述，采取融合域自适应网络和多尺度特征聚合的息肉分割网络可以有效分割息肉图像，并具有良好的泛化性能。

关键词: 息肉分割, 域自适应, 多尺度特征聚合, 医学图像处理