Scholarlink

<p>Accurate pancreatic segmentation in computed tomography (CT) remains challenging due to irregular morphology, low contrast, and blurred boundaries. Most segmentation backbones produce discrete, fixed-grid voxel predictions, which can lead to coarse contours and limited representation of subtle, continuous anatomical deformations. To mitigate this limitation, IRM-UNet is introduced as a backbone-agnostic refinement framework that augments segmentation backbones with an implicit neural representation (INR) module for sub-voxel boundary refinement. The main innovations are as follows: (1) <b>Continuous boundary reconstruction</b>: modeling pancreatic morphology in a continuous domain via an INR enables accurate delineation of complex boundaries; (2) <b>Multi-scale feature fusion and uncertainty-aware sampling</b>: multi-scale feature integration with uncertainty-aware sampling concentrates refinement on low-contrast and ambiguous regions, improving robustness; and (3) <b>High-frequency detail enhancement</b>: combining positional encoding with Fourier feature mapping projects low-dimensional coordinates into a higher-dimensional space, facilitating the representation of high-frequency boundary details. Experiments on the NIH pancreatic CT dataset yield a Dice coefficient of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(89.87 \pm 1.82\%\)</EquationSource> </InlineEquation> and a Jaccard index of <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(81.60 \pm 1.65\%\)</EquationSource> </InlineEquation>. Additional experiments indicate that the implicit refinement module functions as a plug-and-play component across different backbones and is associated with improved boundary quality.</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

IRM-UNet: backbone-agnostic implicit refinement for enhanced pancreatic CT segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月17日

查看全文
  • Chaohui Zhang,
  • Anusha Achuthan

摘要

Accurate pancreatic segmentation in computed tomography (CT) remains challenging due to irregular morphology, low contrast, and blurred boundaries. Most segmentation backbones produce discrete, fixed-grid voxel predictions, which can lead to coarse contours and limited representation of subtle, continuous anatomical deformations. To mitigate this limitation, IRM-UNet is introduced as a backbone-agnostic refinement framework that augments segmentation backbones with an implicit neural representation (INR) module for sub-voxel boundary refinement. The main innovations are as follows: (1) Continuous boundary reconstruction: modeling pancreatic morphology in a continuous domain via an INR enables accurate delineation of complex boundaries; (2) Multi-scale feature fusion and uncertainty-aware sampling: multi-scale feature integration with uncertainty-aware sampling concentrates refinement on low-contrast and ambiguous regions, improving robustness; and (3) High-frequency detail enhancement: combining positional encoding with Fourier feature mapping projects low-dimensional coordinates into a higher-dimensional space, facilitating the representation of high-frequency boundary details. Experiments on the NIH pancreatic CT dataset yield a Dice coefficient of \(89.87 \pm 1.82\%\) and a Jaccard index of \(81.60 \pm 1.65\%\) . Additional experiments indicate that the implicit refinement module functions as a plug-and-play component across different backbones and is associated with improved boundary quality.