Speaker
摘要
半监督医学图像分割（SSMIS）在减少对稀缺标注医学数据依赖方面展现出良好前景。然而，该领域仍面临诸多挑战，例如标注数据与无标注数据之间存在分布不匹配、现有图像混合策略（如随机裁剪）极易破坏医学解剖结构的连贯性，以及混合过程中引入大量无关背景噪声等问题。为解决这些难题，本文提出了一种显著性引导的半监督医学图像分割框架（SC-BCP）。该框架利用模型自身的梯度反馈，设计了一种显著性引导掩码（Saliency-Guided Mask, SGM）生成机制。具体而言，该机制首先提取置信度加权的前景梯度响应，随后通过形态学提纯与焦点区域评估，动态锁定最具信息量的解剖结构；最后提取其宏观特征生成可控的几何混合掩码。这一策略替代了传统的随机矩形掩码，在图像混合中有效保护了器官边界的完整性并抑制了背景干扰。在 ACDC 和 PROMISE12 数据集上的实验验证了该框架的有效性。相对于基线BCP方法，SC-BCP 取得了显著的性能提升，在ACDC数据集，仅使用10% 标注数据下，Dice系数提升了1.35%，95HD下降了2.4mm；在PROMISE12数据集，使用20% 标注数据下Dice系数提升了15.71%，95HD下降了9.63mm。
Abstract
Semi-supervised medical image segmentation (SSMIS) has shown great promise in reducing reliance on scarce annotated medical data. However, it still faces several challenges, including the distribution mismatch between labeled and unlabeled data, the disruption of anatomical structure caused by conventional image mixing strategies such as random cropping, and the introduction of excessive irrelevant background noise during the mixing process. To address these issues, we propose a saliency-guided semi-supervised medical image segmentation framework, termed SC-BCP. The proposed framework leverages the model's own gradient feedback to develop a Saliency-Guided Mask (SGM) generation mechanism. Specifically, it first extracts confidence-weighted foreground gradient responses, then dynamically identifies the most informative anatomical structures through morphological refinement and focal-region assessment, and finally derives controllable geometric mixing masks from their macroscopic features. Compared with conventional random rectangular masks, the proposed strategy better preserves organ boundary integrity and suppresses background interference during image mixing. Experiments on the ACDC and PROMISE12 datasets demonstrate the effectiveness of SC-BCP. Compared with the baseline BCP method, SC-BCP achieves significant performance improvements. On ACDC, using only 10% labeled data, it improves the Dice coefficient by 1.35% and reduces the 95% Hausdorff Distance (95HD) by 2.4 mm. On PROMISE12, with 20% labeled data, it improves the Dice coefficient by 15.71% and reduces the 95HD by 9.63 mm.
| 关键词 | 医学图像分割；半监督学习；一致性正则化；显著图 |
|---|---|
| Keywords | Medical image segmentation；Semi-supervised learning； Consistency regularizationy；Saliency maps |
