| 摘 要: 针对肝脏肿瘤分割中小目标丢失、边界不规则及背景噪声干扰等问题，提出了一种融合先验引导、形态感知与门控机制的分割方法。该方法引入先验引导注意力模块(PGA)利用先验图强化病灶特征；构建形态特征提取模块(MFE)通过形态重构优化复杂边界细节；设计实例级门控模块(IGM)建立自动筛选机制以抑制噪声。LiTS数据集实验表明，该方法优于对比的主流网络，相较U-Net基线，肿瘤Dice与IoU分别提升8.87与9.75个百分点，验证了其在兼顾器官结构一致性与病灶精细分割上的有效性。 |
| 关键词: 肝脏肿瘤分割 先验引导 形态感知 门控机制 |
|
中图分类号:
文献标识码:
|
| 基金项目: 国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| Prior-guided Morphology-aware Gated Method for Liver Tumor Segmentation |
|
Wu Jian Qiang1, Wang Weidong2, Qin Bin2
|
1.JiangSu University Of Science And Technology;2.Jiangsu University of Science and Technology
|
| Abstract: To address challenges such as the loss of small tumor information, irregular boundaries, and background noise interference in liver tumor segmentation, a segmentation method integrating prior guidance, morphology awareness, and gating mechanisms is proposed. This method introduces a Prior-Guided Attention (PGA) module that utilizes prior maps to enhance lesion features; constructs a Morphology Feature Extractor (MFE) to optimize complex boundary details through morphological reconstruction; and designs an Instance-wise Gating Module (IGM) to establish an automatic screening mechanism for noise suppression. Experiments on the LiTS dataset demonstrate that the proposed method outperforms comparative mainstream networks. Compared to the U-Net baseline, the Dice and IoU for tumors improved by 8.87 and 9.75 percentage points, respectively, validating its effectiveness in balancing organ structural consistency with precise lesion segmentation. |
| Keywords: Liver tumor segmentation Prior guidance Morphology awareness Gating mechanism |
