Scholarlink

<p>Accurate liver and tumor segmentation from CT images is essential for cancer diagnosis, treatment planning, and response assessment. However, manual segmentation is labor-intensive and variable, while standard automated models lack the flexibility to adapt to diverse clinical needs or inherent image uncertainties. To bridge this gap, we introduce User-Preference Alignment with Uncertainty-Aware Interactive Rectification (UAIR), a novel framework designed for efficient and adaptive segmentation. Instead of requiring laborious pixel-level corrections, UAIR presents the clinician with a small, curated set of diverse segmentation candidates generated by quantifying model uncertainty. The user simply selects the most suitable option, allowing the framework to iteratively refine its results and align with specific clinical preferences. This selection-based approach drastically reduces the human interaction cost. We validated UAIR on a large-scale, multi-center CT dataset, demonstrating superior accuracy (DSC 0.776) over existing manual positional prompting (DSC 0.685) and less prompting efforts. UAIR provides a clinically-viable solution that integrates seamless human guidance, enabling rapid and robust segmentation for downstream quantitative analysis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

User-preference alignment with uncertainty-aware interactive rectification for liver organ and tumor segmentation and analysis from CT images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月3日

查看全文
  • Guangyuan Zhao,
  • Yang Wang,
  • Chen Gong,
  • Zipei Wang,
  • Guobin Huang,
  • Xuechun Zhao,
  • Tengfei Chao,
  • Bo Yang

摘要

Accurate liver and tumor segmentation from CT images is essential for cancer diagnosis, treatment planning, and response assessment. However, manual segmentation is labor-intensive and variable, while standard automated models lack the flexibility to adapt to diverse clinical needs or inherent image uncertainties. To bridge this gap, we introduce User-Preference Alignment with Uncertainty-Aware Interactive Rectification (UAIR), a novel framework designed for efficient and adaptive segmentation. Instead of requiring laborious pixel-level corrections, UAIR presents the clinician with a small, curated set of diverse segmentation candidates generated by quantifying model uncertainty. The user simply selects the most suitable option, allowing the framework to iteratively refine its results and align with specific clinical preferences. This selection-based approach drastically reduces the human interaction cost. We validated UAIR on a large-scale, multi-center CT dataset, demonstrating superior accuracy (DSC 0.776) over existing manual positional prompting (DSC 0.685) and less prompting efforts. UAIR provides a clinically-viable solution that integrates seamless human guidance, enabling rapid and robust segmentation for downstream quantitative analysis.