Song Yuhui, Xu Chenchu, Wang Boyan, Du Xiuquan, Chen Jie, Zhang Yanping, Li Shuo
School of Computer Science and Technology, Anhui University, 230601, Hefei, China; Key Laboratory of Intelligent Computing and Signal Processing of Ministry of Education, Anhui University, 230601, Hefei, China.
School of Computer Science and Technology, Anhui University, 230601, Hefei, China; Key Laboratory of Intelligent Computing and Signal Processing of Ministry of Education, Anhui University, 230601, Hefei, China; Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, 230051, Hefei, China.
Artif Intell Med. 2025 Sep;167:103183. doi: 10.1016/j.artmed.2025.103183. Epub 2025 Jun 18.
Few-shot learning alleviates the heavy dependence of medical image segmentation on large-scale labeled data, but it shows strong performance gaps when dealing with new tasks compared with traditional deep learning. Existing methods mainly learn the class knowledge of a few known (support) samples and extend it to unknown (query) samples. However, the large distribution differences between the support image and the query image lead to serious deviations in the transfer of class knowledge, which can be specifically summarized as two segmentation challenges: Intra-class inconsistency and Inter-class similarity, blurred and confused boundaries. In this paper, we propose a new interactive prototype learning and self-learning network to solve the above challenges. First, we propose a deep encoding-decoding module to learn the high-level features of the support and query images to build peak prototypes with the greatest semantic information and provide semantic guidance for segmentation. Then, we propose an interactive prototype learning module to improve intra-class feature consistency and reduce inter-class feature similarity by conducting mid-level features-based mean prototype interaction and high-level features-based peak prototype interaction. Last, we propose a query features-guided self-learning module to separate foreground and background at the feature level and combine low-level feature maps to complement boundary information. Our model achieves competitive segmentation performance on benchmark datasets and shows substantial improvement in generalization ability.
少样本学习减轻了医学图像分割对大规模标注数据的严重依赖,但与传统深度学习相比,在处理新任务时表现出较大的性能差距。现有方法主要学习少数已知(支持)样本的类别知识,并将其扩展到未知(查询)样本。然而,支持图像和查询图像之间的巨大分布差异导致类别知识转移时出现严重偏差,具体可归纳为两个分割挑战:类内不一致和类间相似性,边界模糊且混淆。在本文中,我们提出了一种新的交互式原型学习与自学习网络来解决上述挑战。首先,我们提出一个深度编码-解码模块来学习支持图像和查询图像的高级特征,以构建具有最大语义信息的峰值原型,并为分割提供语义指导。然后,我们提出一个交互式原型学习模块,通过基于中级特征的均值原型交互和基于高级特征的峰值原型交互来提高类内特征一致性并降低类间特征相似性。最后,我们提出一个查询特征引导的自学习模块,在特征层面分离前景和背景,并结合低级特征图来补充边界信息。我们的模型在基准数据集上取得了有竞争力的分割性能,并在泛化能力方面有显著提升。