Ni Lina, Liu Yang, Zhang Zekun, Li Yongtao, Zhang Jinquan
College of Computer Science and Engineering, Shandong University of Science and Technology, Qingdao 266590, China.
Key Laboratory of the Ministry of Education for Embedded System and Service Computing, Tongji University, Shanghai 201804, China.
Sensors (Basel). 2025 Mar 29;25(7):2176. doi: 10.3390/s25072176.
Few-shot learning has demonstrated remarkable performance in medical image segmentation. However, existing few-shot medical image segmentation (FSMIS) models often struggle to fully utilize query image information, leading to prototype bias and limited generalization ability. To address these issues, we propose the dual-filter cross attention and onion pooling network (DCOP-Net) for FSMIS. DCOP-Net consists of a prototype learning stage and a segmentation stage. During the prototype learning stage, we introduce a dual-filter cross attention (DFCA) module to avoid entanglement between query background features and support foreground features, effectively integrating query foreground features into support prototypes. Additionally, we design an onion pooling (OP) module that combines eroding mask operations with masked average pooling to generate multiple prototypes, preserving contextual information and mitigating prototype bias. In the segmentation stage, we present a parallel threshold perception (PTP) module to generate robust thresholds for foreground and background differentiation and a query self-reference regularization (QSR) strategy to enhance model accuracy and consistency. Extensive experiments on three publicly available medical image datasets demonstrate that DCOP-Net outperforms state-of-the-art methods, exhibiting superior segmentation and generalization capabilities.
少样本学习在医学图像分割中已展现出卓越的性能。然而，现有的少样本医学图像分割（FSMIS）模型常常难以充分利用查询图像信息，导致原型偏差和泛化能力受限。为解决这些问题，我们提出了用于FSMIS的双滤波器交叉注意力与洋葱池化网络（DCOP-Net）。DCOP-Net由一个原型学习阶段和一个分割阶段组成。在原型学习阶段，我们引入双滤波器交叉注意力（DFCA）模块，以避免查询背景特征与支持前景特征之间的纠缠，有效地将查询前景特征整合到支持原型中。此外，我们设计了一个洋葱池化（OP）模块，该模块将腐蚀掩码操作与掩码平均池化相结合以生成多个原型，保留上下文信息并减轻原型偏差。在分割阶段，我们提出了一个并行阈值感知（PTP）模块，用于生成用于前景和背景区分的鲁棒阈值，以及一种查询自参考正则化（QSR）策略，以提高模型的准确性和一致性。在三个公开可用的医学图像数据集上进行的大量实验表明，DCOP-Net优于现有方法，展现出卓越的分割和泛化能力。
