Skip to main content
QUICK REVIEW

[论文解读] NeuroSymb-MRG: Differentiable Abductive Reasoning with Active Uncertainty Minimization for Radiology Report Generation

Rong Fu, Yiqing Lyu|arXiv (Cornell University)|Mar 2, 2026
Topic Modeling被引用 0
一句话总结

NeuroSymb-MRG 将可微分的神经符号组合理性推理与检索增强生成以及主动不确定性最小化相结合,生成结构化、临床 grounding 的放射科报告,提升事实性和标准评价指标。

ABSTRACT

Automatic generation of radiology reports seeks to reduce clinician workload while improving documentation consistency. Existing methods that adopt encoder-decoder or retrieval-augmented pipelines achieve progress in fluency but remain vulnerable to visual-linguistic biases, factual inconsistency, and lack of explicit multi-hop clinical reasoning. We present NeuroSymb-MRG, a unified framework that integrates NeuroSymbolic abductive reasoning with active uncertainty minimization to produce structured, clinically grounded reports. The system maps image features to probabilistic clinical concepts, composes differentiable logic-based reasoning chains, decodes those chains into templated clauses, and refines the textual output via retrieval and constrained language-model editing. An active sampling loop driven by rule-level uncertainty and diversity guides clinician-in-the-loop adjudication and promptbook refinement. Experiments on standard benchmarks demonstrate consistent improvements in factual consistency and standard language metrics compared to representative baselines.

研究动机与目标

  • 推动自动放射科报告生成,使其 fluent 但在事实性上可靠且临床可解释。
  • 开发一个统一框架,将图像映射到概率临床概念,构建可微分的推理链,并生成模板化文本输出。
  • 结合检索证据和受约束的语言模型改写来润色草稿。
  • 引入主动不确定性最小化,以便优先让临床医生审核高价值、高不确定性案例。
  • 在代表性基线的标准放射科报告基准上展示改进。

提出的方法

  • 通过基于 transformer 的概念预测器将图像特征映射到概率临床概念。
  • 在概念上构建可微分的软逻辑推理链,使用具备 AND/OR/NOT 运算符的可微分逻辑层。
  • 将激活的软规则解码为规范子句模板,使用规则引导的解码器。
  • 在解码子句上增加检索证据与受约束的 LLM 复述,形成结构化草稿。
  • 使用多主体编排系统(知识、验证、推理)结合检索增强的模板填充管线。
  • 通过基于蒙特卡洛 dropout 的熵与 k-center 多样性采样引导临床医生协同 refinement,实施主动不确定性最小化。
Figure 1: Architectural overview of the NeuroSymb-MRG framework for transparent and clinically grounded radiology report generation. The pipeline initiates with Visual Perception , utilizing a self-supervised visual encoder $f_{\mathrm{ve}}$ to extract patch-level features $X$ . In the Neuro-Symboli
Figure 1: Architectural overview of the NeuroSymb-MRG framework for transparent and clinically grounded radiology report generation. The pipeline initiates with Visual Perception , utilizing a self-supervised visual encoder $f_{\mathrm{ve}}$ to extract patch-level features $X$ . In the Neuro-Symboli

实验结果

研究问题

  • RQ1可微分神经符号推理模块是否比纯神经基线产生更一致事实性的放射科报告?
  • RQ2在规则层面的主动不确定性最小化是否提升事实可靠性并减少临床风险的幻觉?
  • RQ3当与基于规则的解码器相结合时,检索增强与受约束的 LLM 改写对报告质量有何影响?
  • RQ4具知识图谱(如 UMLS)的一种多主体编排是否能提升报告的真实度与矛盾处理能力?

主要发现

MethodB-1 IU X-rayB-2 IU X-rayB-3 IU X-rayB-4 IU X-rayR-L IU X-rayMTR IU X-rayB-1 MIMIC-CXRB-2 MIMIC-CXRB-3 MIMIC-CXRB-4 MIMIC-CXRR-L MIMIC-CXRMTR MIMIC-CXR
Show-Tell [27]0.2430.1300.1080.0780.3070.1570.3080.1900.1250.0880.2560.122
Transformer [26]0.3720.2510.1470.1360.3170.1680.3160.1990.1400.0920.2670.129
Att2in [24]0.2480.1340.1160.0910.3090.1620.3140.1980.1330.0950.2640.122
AdaAtt [20]0.2840.2070.1500.1260.3110.1650.3140.1980.1320.0940.2670.128
Up-Down [2]0.3170.1950.1300.0920.2670.128
M2Transformer [6]0.4020.2840.1680.1430.3280.1700.3320.2100.1420.1010.2640.134
R2Gen [5]0.4700.3040.2190.1650.3710.1870.3530.2180.1450.1030.2770.142
Contra.Attn. [19]0.4920.3140.2220.1690.3810.1930.3500.2190.1520.1090.2830.151
CMCL [17]0.4730.3050.2170.1620.3780.1860.3440.2170.1400.0970.2810.133
CMN [4]0.4750.3090.2220.1700.3750.1910.3530.2180.1480.1060.2780.142
Aligntransformer [32]0.4840.3130.2250.1730.3790.2040.3780.2350.1560.1120.2830.158
M2Tr.Prog. [21]0.4860.3170.2320.1730.3900.1920.3780.2320.1540.1070.2720.145
CMM+RL [22]0.4810.3160.2280.1810.3840.2010.3810.2320.1550.1090.2870.151
XPRONET* [28]0.4910.3250.2280.1690.3870.2020.3440.2150.1460.1050.2790.138
MCGN [30]0.4810.3160.2260.1710.3720.1900.3730.2350.1620.1200.2820.143
PPKED [18]0.4830.3150.2240.1680.3760.3600.2240.1490.1060.2840.149
RAMT [33]0.4820.3100.2210.1650.3770.1950.3620.2290.1570.1130.2840.153
R2GenGPT [29]0.4820.3060.2150.1580.3700.2000.3870.2480.1700.1230.2800.149
VLCI ${\dagger}$ [3]0.3240.2110.1510.1150.3790.1660.3570.2160.1440.1030.2560.136
PromptMRG [11]0.4010.2810.1600.3980.1120.2680.157
MedRAT [9]0.4550.3490.3650.0860.251
MRG-LLM [14]0.5290.3590.2660.2020.4080.2210.4160.2670.1820.1290.2960.163
NeuroSymb-MRG (Ours)0.6020.4250.3210.2530.4630.2750.4870.3320.2340.1750.3620.225
  • NeuroSymb-MRG 在 IU X-ray 与 MIMIC-CXR 上的自动化指标高于强基线(BLEU、ROUGE-L、METEOR)。
  • 基于规则的可微推理结合规则解码器在词汇与语义质量上优于仅 MLP 或不含符号推理的基线。
  • 检索增强与 LLM 约束步骤带来收益,并通过验证器帮助减轻矛盾。
  • 通过熵基采样与多样性(k-center)实现的主动不确定性最小化在保持质量的同时降低标注需求。
  • 具知识代理与 UMLS 验证的多主体设置提升事实性合理性并降低矛盾陈述。

更好的研究,从现在开始

从论文设计到论文写作,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。