文章背景与核心概要

放射学报告通常包含非结构化的临床观察结果，涵盖现病史、否定病史、既往史、不确定或无关的发现，并且往往伴随多个共存的异常。传统的分割方法通过依赖预先指定的目标标识或空间提示（充当隐藏的目标神谕）来规避这种复杂性。

为了解决这一局限性，作者推出了 EliSeg，这是一种专为基于报告的异常分割（report-grounded abnormality segmentation）设计的新颖框架。EliSeg 不依赖手动空间提示，而是直接从未经筛选的报告中确定目标合格性、基数（cardinality）以及发现到掩码的对应关系，从而清晰地分割出相应的区域。

EliSeg: Verified Target Construction for Report-Grounded Abnormality Segmentation

arXiv ID: arXiv:2608.07299
Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
Submitted: August 7, 2026
Authors: Chengyi Peng, Haoyu Yang, Meixing Shi, Yuxiang Cai, Yankai Jiang
Links: View PDF | GitHub Repository

📌 Summary

放射学报告经常包含非结构化的临床观察结果，涵盖当前、否定、既往、不确定或不相关的发现，并且通常伴随着多个共存的异常。传统的分割方法通过依赖预先指定的目标身份或空间提示（充当隐藏的目标神谕）来绕过这种复杂性。

Radiology reports frequently contain unstructured clinical observations encompassing present, negated, prior, uncertain, or irrelevant findings, often alongside multiple co-existing abnormalities. Traditional segmentation methods sidestep this complexity by relying on pre-specified target identities or spatial prompts (acting as a hidden target oracle).

为了解决这一局限性，作者引入了 EliSeg，这是一个专为基于报告的异常分割而设计的新颖框架。与依赖手动空间提示不同，EliSeg 直接从过滤后的报告中确定目标的合格性、基数以及发现与掩码的对应关系，从而干净利落地分割出相应的区域。

To address this limitation, the authors introduce EliSeg, a novel framework designed for report-grounded abnormality segmentation. Instead of relying on manual spatial prompts, EliSeg determines target eligibility, cardinality, and finding-to-mask correspondence directly from unfiltered reports, cleanly segmenting the corresponding regions.

🛠️ Methodology: The Actor–Verify–Revise Framework

EliSeg 通过三步机制将目标构建与掩码生成结合起来：

EliSeg integrates target construction with mask generation through a three-step mechanism:

1. 行动者（语法约束）： 根据报告提出初始目标槽和相应的分割掩码。

   Actor (Grammar-Constrained): Proposes initial target slots and corresponding segmentation masks based on the report.

2. 验证器（纯文本）： 一个独立的纯文本模块，用于重建符合条件的发现清单，以验证所提出的目标。

   Verifier (Text-Only): An independent text-only module that reconstructs the eligible finding inventory to validate the proposed targets.

3. 修正： 每当所提出的目标结构与验证的发现之间出现差异时，有选择地重新执行共享的行动者。

   Revision: Selectively re-executes the shared Actor whenever discrepancies arise between the proposed target structures and the verified findings.

注：EliSeg 的运行完全不需要预定义的目标身份、发现提示、点标记或边界框。

Note: EliSeg functions entirely without predefined target identities, finding prompts, point markers, or bounding boxes.

📊 Results & Performance

• MIMIC-CXR-ILS： 在各种医学发现中，EliSeg 的表现始终优于标准的直接分割方法和传统的“先提取后分割”级联方法。它还成功抑制了为不符合条件或被否定的报告提及生成的掩码。

  • MIMIC-CXR-ILS: EliSeg consistently outperforms standard direct segmentation methods and traditional extract-then-segment cascades across various medical findings. It also successfully suppresses masks generated for ineligible or negated report mentions.

• 消融实验： 证实了验证和修正组件在提高整体准确性方面起到的关键且互补的作用。

  • Ablation Studies: Confirmed the crucial and complementary roles of both the verification and revision components in improving overall accuracy.

• CheXlocalize： 展示了对外部环境强大的跨数据集迁移能力。

  • CheXlocalize: Demonstrates robust cross-dataset transferability to external environments.