%0 Journal Article
%T 面向深度感知的空间层次学习用于稳健的医学图像分割
Depth-Aware Spatial Hierarchy Learning for Robust Medical Image Segmentation
%A 周世钦
%J Software Engineering and Applications
%P 494-509
%@ 2325-2278
%D 2026
%I Hans Publishing
%R 10.12677/sea.2026.153046
%X 皮肤镜图像分割是计算机辅助诊断皮肤癌的关键步骤，但现有方法多依赖二维纹理信息，忽略了单目图像中隐含的视觉远近线索(伪深度)。本文提出一种融合伪深度注意力机制与边界监督的多任务U-Net框架——DASH-Net (Depth-Aware Spatial Hierarchy Network)。其核心创新在于通过深度引导的空间–通道双注意力模块(DSCA)，建立二维图像特征通道与视觉远近层次之间的自适应关联，使分割网络能够根据深度先验选择性地增强来自相关深度范围的特征通道。此外，针对临床图像提出轻量变体DASH-Light，以亮度映射替代伪深度图，实现无外部深度估计模型的高效分割。在ISIC 2018皮肤病变基准数据集上的实验表明，DASH-Net的Dice系数达88.99%，Jaccard指数达82.02%，参数量仅7.85 M；替换为PVTv2骨干后Dice进一步提升至90.61%。在私有静脉畸形临床数据集上，DASH-Light取得66.80% Dice和30.21的HD95，较当前最优方法提升1.1%和下降12.4。消融实验与可视化分析揭示了DSCA模块的工作机制：部分通道专门化编码深层病灶信息，部分编码背景表层信息，形成类似人类视觉系统的深度语义分工。
Skin lesion segmentation is a critical step in computer-aided diagnosis of skin cancer. Existing methods mostly rely on two-dimensional texture information and ignore the implicit visual depth cues in monocular images. This study proposes DASH-Net, a multi-task U-Net framework integrating pseudo-depth attention with boundary supervision. Its core innovation is the Depth-guided Spatial-Channel Attention (DSCA) module that establishes an adaptive association between image feature channels and visual depth hierarchy. A lightweight variant DASH-Light is also proposed for clinical images where brightness serves as a depth proxy. On ISIC 2018, DASH-Net achieves 88.99% Dice with only 7.85 M parameters, and reaches 90.61% Dice with PVTv2 backbone. On a private venous malformation clinical dataset, DASH-Light achieved a 66.80% Dice score and a 30.21 HD95, improving by 1.1% and decreasing by 12.4 compared to the current best method. Ablation studies reveal that DSCA channels spontaneously specialize into deep-preferring and background-preferring groups, forming a depth semantic division.
%K 皮肤病变分割，
%K 多任务学习，
%K 伪深度估计，
%K 空间–
%K 通道注意力，
%K 边界监督，
%K 静脉畸形
Skin Lesion Segmentation
%K Multi-Task Learning
%K Pseudo-Depth Estimation
%K Spatial-Channel Attention
%K Boundary Supervision
%K Venous Malformation
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=145452