基于双视角注意力的小样本医学图像分割

Few-Shot Medical Image Segmentation Based on Dual-Perspective Attention

摘要: 专业医学图像的采集成本与标注要求高，导致可用标注数据量不足，小样本医学图像分割成为当前解决该问题的重要方向。传统的方法通常从支撑图像中学习原型，并通过最近邻搜索的方法分割查询图像。尽管这些算法展现出较高的性能，但由于部分容积效应导致组织边界模糊；灰度不均匀性造成同类组织表征差异；前景和背景共享相似视觉特征，使得模型易产生混杂错误分割。为此本文提出双视角协同增强网络(Dual-Perspective Collaborative Enhancement Network, DPCENet)，通过空间–通道双视角注意力机制实现协同精准特征筛选，具体来说，空间注意力精确定位器官空间分布；通道注意力动态增强相关语义通道，抑制无关噪声。考虑到医学图像前景和背景共享许多视觉特征，模型采用双路径平衡设计，并行处理前景与背景特征流。在三个公开的医学图像数据集上的实验表明，该方法的性能优于现有方法，详细的分析也验证了该模块的有效性。

Abstract: The high acquisition cost and high annotation requirements of professional medical images lead to insufficient available annotated data. Small sample medical image segmentation has become an important direction to solve this problem. Traditional methods usually learn prototypes from supporting images and segment query images through nearest neighbor search. Although these algorithms show high performance, the partial volume effect leads to blurred tissue boundaries; grayscale inhomogeneity causes differences in the representation of similar tissues; the foreground and background share similar visual features, which makes the model prone to mixed mis-segmentation. To this end, this paper proposes a dual-perspective collaborative enhancement network (DPCENet), which achieves collaborative and precise feature screening through a spatial-channel dual-perspective attention mechanism. Specifically, spatial attention accurately locates the spatial distribution of organs; channel attention dynamically enhances related semantic channels and suppresses irrelevant noise. Considering that the foreground and background of medical images share many visual features, the model adopts a dual-path balanced design to process the foreground and background feature streams in parallel. Experiments on three public medical image datasets show that the performance of this method is better than that of existing methods, and detailed analysis also verifies the effectiveness of this module.