空间–频率域分层特征增强的乳腺超声图像分割

Spatial-Frequency Domain Hierarchical Feature Enhancement for Breast Ultrasound Image Segmentation

摘要: 乳腺超声图像分割是计算机辅助诊断系统的关键环节，但斑点噪声干扰与病灶边界模糊给精确分割带来了严峻挑战。现有方法多采用统一策略处理跳跃连接中的多层级特征，忽视了浅层特征噪声敏感与深层特征定位模糊的本质差异。针对上述问题，提出一种空间–频率域分层特征增强的轻量化分割网络SFH-Net (Spatial-Frequency Hierarchical Network)。首先，设计交替双路径编码器，通过标准卷积与深度可分离卷积的交替组合并引入空洞卷积扩展感受野，在保证特征提取能力的同时有效控制参数规模；其次，提出分层跳跃连接策略，对浅层特征引入自适应频域滤波模块以抑制斑点噪声，对深层特征采用坐标注意力机制增强空间定位能力。在BUSI和UDIAT两个公开数据集上的实验表明，SFH-Net以仅13.12 M的参数量分别取得了78.73%和85.15%的Dice系数，所提出方法能够在保持较低模型复杂度的同时实现有效的病灶分割。

Abstract: Breast ultrasound image segmentation is a crucial component of computer-aided diagnosis systems, yet speckle noise interference and blurred lesion boundaries pose severe challenges to accurate segmentation. Existing methods mostly adopt a unified strategy to process multi-level features in skip connections, ignoring the essential differences between shallow features that are sensitive to noise and deep features with imprecise localization, and also underutilize frequency domain information. To address these issues, we propose a lightweight segmentation network named SFH-Net with spatial-frequency domain hierarchical feature enhancement. First, an alternating dual-path encoder is designed: the alternating combination of standard convolution and depthwise separable convolution, combined with the introduction of atrous convolution to expand the receptive field, effectively controls the parameter scale while ensuring the feature extraction capability. Second, a hierarchical skip connection strategy is proposed: an adaptive frequency domain filtering module is introduced for shallow features to suppress speckle noise, and the coordinate attention mechanism is adopted for deep features to enhance spatial localization ability. Experiments on two public datasets (BUSI and UDIAT) demonstrate that SFH-Net achieves Dice coefficients of 78.73% and 85.15% with only 13.12 million parameters, the proposed method can achieve effective lesion segmentation while maintaining low model complexity.