NeuroSymb-MRG: Differentiable Abductive Reasoning with Active Uncertainty Minimization for Radiology Report Generation

Rong Fu; Yiqing Lyu; Chunlei Meng; Muge Qi; Yabin Jin; Qi Zhao; Li Bao; Juntao Gao

cs.CVMar 2, 2026

NeuroSymb-MRG: Differentiable Abductive Reasoning with Active Uncertainty Minimization for Radiology Report Generation

Authors: Rong Fu, Yiqing Lyu, Chunlei Meng, Muge Qi, Yabin Jin, Qi Zhao, Li Bao, Juntao Gao, +4 more

Organizations: University of Macau · Tsinghua University · Fudan University · Peking University · The First People’s Hospital of Foshan · Capital Medical University · Rutgers Cancer Institute, NJ, USA · NYU Langone Health, NY, USA · Techno International New Town, Kolkata, India

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.

Figures & tables

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-Symbolic Reasoning module, these features are mapped to probabilistic concept activations $\hat{c}$ , which serve as leaves for a Differentiable Logic Layer that composes multi-hop reasoning chains through soft logical operators (AND, OR, NOT). The Multi-Agent Orchestration hub employs a request-response protocol among specialized agents (Knowledge, Verifier, Reasoning) to cross-reference findings against a Medical Knowledge Graph (UMLS) . The Hybrid Generation phase merges rule-decoded clauses $\widehat{\mathcal{L}}$ with retrieval-augmented evidence $\mathcal{R}$ via a template inventory $\mathcal{T}$ and constrained LLM refinement. Finally, an Active Uncertainty Minimization loop utilizes predictive entropy and a Feedback Simulator to identify high-value cases for clinician adjudication, iteratively refining the promptbook and model parameters.

Method	IU X-ray [ 12 ]						MIMIC-CXR [ 11 ]
Method	B-1	B-2	B-3	B-4	R-L	MTR	B-1	B-2	B-3	B-4	R-L	MTR
Show-Tell [ 1 ]	0.243	0.130	0.108	0.078	0.307	0.157	0.308	0.190	0.125	0.088	0.256	0.122
Transformer [ 2 ]	0.372	0.251	0.147	0.136	0.317	0.168	0.316	0.199	0.140	0.092	0.267	0.129
Att2in [ 31 ]	0.248	0.134	0.116	0.091	0.309	0.162	0.314	0.198	0.133	0.095	0.264	0.122
AdaAtt [ 3 ]	0.284	0.207	0.150	0.126	0.311	0.165	0.314	0.198	0.132	0.094	0.267	0.128
Up-Down [ 4 ]	–	–	–	–	–	–	0.317	0.195	0.130	0.092	0.267	0.128

Table 1: Quantitative comparison with representative prior methods on IU X-ray and MIMIC-CXR. Best and second-best values are in bold and underlined , respectively.

Model Variant	BLEU-1	BLEU-2	BLEU-3	BLEU-4	ROUGE-L	METEOR	$\Delta$ B-1
Full NeuroSymb-MRG	0.487	0.332	0.234	0.175	0.362	0.225	–
I. Core Architectural Components
w/o Self-Supervised Visual Encoder (Random Init)	0.412	0.262	0.175	0.124	0.298	0.168	-15.4%
w/o Concept Attention Head (Raw Patch Features)	0.401	0.251	0.168	0.118	0.288	0.159	-17.7%
w/o Differentiable Logic Layer (End-to-End MLP)	0.428	0.278	0.190	0.136	0.315	0.185	-12.1%
w/o Rule-Guided Decoder (Direct LLM Generation)	0.439	0.289	0.198	0.144	0.321	0.191	-9.9%

Table 2: Comprehensive ablation on architectural components and key mechanisms of NeuroSymb-MRG evaluated on MIMIC-CXR. The symbol

\downarrow

indicates performance drop relative to the full model.

Training & Retrieval Strategy	BLEU-1	BLEU-2	BLEU-3	BLEU-4	ROUGE-L	METEOR
Full Strategy	0.487	0.332	0.234	0.175	0.362	0.225
Loss weighting: Fixed weights	0.471	0.318	0.222	0.165	0.348	0.213
Loss weighting: GradNorm	0.480	0.327	0.230	0.172	0.357	0.221
No rule supervision ( $\mathcal{L}_{\mathrm{rule}}$ removed)	0.435	0.285	0.197	0.145	0.322	0.192
Active sampling: Random selection	0.463	0.310	0.217	0.161	0.342	0.208
Active sampling: Entropy only	0.475	0.323	0.228	0.170	0.353	0.217

Table 3: Ablation on training strategies and retrieval settings. Dynamic loss weighting and combined active sampling are most effective.

Logical Reasoning Configuration	BLEU-1	BLEU-2	BLEU-3	BLEU-4	ROUGE-L	METEOR
Full: Soft Tree (Prod. T-Norm + Prob. Sum + Gate)	0.487	0.332	0.234	0.175	0.362	0.225
Operator ablation: Min/Max T-norm / Conorm	0.479	0.325	0.228	0.170	0.355	0.219
Operator ablation: Łukasiewicz T-norm / Conorm	0.473	0.319	0.223	0.166	0.350	0.214
Operator ablation: Gödel T-norm / Conorm	0.468	0.315	0.219	0.162	0.345	0.210
Composition ablation: Fixed $\alpha=1.0$ (AND only)	0.452	0.299	0.206	0.151	0.332	0.200
Composition ablation: Fixed $\alpha=0.0$ (OR only)	0.461	0.308	0.214	0.158	0.340	0.207

Table 4: Ablation on the differentiable logic module and decoding strategies evaluated on MIMIC-CXR.

Explore similar work

UR $^{2}$ -MLLM: Uncertainty-aware Revisit Reasoning in Multimodal Large Language Models for Radiology Report Generation
Aug 23, 2026Yucheng Chen, Yang Yu, Jiazhou Zhou +5Radiology Report Generation Radiology VLMs
MedVisAI Lab · Lee Kong Chian School of Medicine, Nanyang Technological University, Singapore · Centre of AI in Medicine, Singapore +3
CliQ-RRG: Clinical-Knowledge Guided Disease-aware Visual-Textual Alignment for QA-Style Radiology Report Generation
Oct 9, 2026Chandravardhan Singh Raghaw, Shahid Shafi Dar, Arnav Nishith Sharma +2Radiology Report Generation Medical VLMs
Indian Institute of Technology Indore, Indore, India · University of Texas at Dallas, USA · University of Wisconsin - Madison, USA
MedScribe: Clinically Grounded CT Reporting through Agentic Workflows
May 3, 2026Giuseppe A. Orlando, Paolo Papotti, Maria A. Zuluaga +2Radiology Report Generation Healthcare
Inria Centre at Université Côte d’Azur Sophia Antipolis, France · EURECOM Biot, France · Dep. of Nuclear Medicine, Centre Antoine Lacassagne Nice, France Université Côte d’Azur, CNRS, Inserm, iBV

Aug 23, 2026J/K move · Enter open · S save

UR $^{2}$ -MLLM: Uncertainty-aware Revisit Reasoning in Multimodal Large Language Models for Radiology Report Generation

Yucheng Chen, Yang Yu, Jiazhou Zhou, Yufei Shi+4

MedVisAI Lab · Lee Kong Chian School of Medicine, Nanyang Technological University, Singapore · Centre of AI in Medicine, Singapore+3

Radiology Report Generation Radiology VLMs Multimodal Large Language Models Multimodal Reasoning

Open paper AskPDF arXiv

Radiologists generate diagnostic reports through iterative and selective revisiting of suspicious regions to refine their interpretations. Recent multimodal large language models (MLLMs) for radiology report generation (RRG) have shifted from text-only reasoning toward a ``Thinking-with-Images'' paradigm, incorporating visual evidence into the reasoning process. However, existing methods provide static visual evidence without a dynamic revisit mechanism during reasoning, neglecting how radiologists re-examine uncertain observations. To this end, we propose an Uncertainty-aware Revisit Reasoning MLLM (UR

^{2}

-MLLM) framework that dynamically revisits uncertain regions during reasoning for RRG. UR

^{2}

-MLLM is first equipped with uncertainty perception by training on an uncertainty-aware dataset. We then construct a multimodal reasoning trajectory dataset together with a detect-and-copy mechanism, which guides when and where to revisit. Finally, a visual grounding reward refines this behavior through reinforcement learning, aligning the revisited regions with corresponding anatomical structures. Experiments on MIMIC-CXR and IU-Xray show that UR

^{2}

-MLLM achieves state-of-the-art performance, highlighting the value of uncertainty-aware visual revisit reasoning for reliable and clinically aligned report generation.

Abstract

Explore similar work

UR2^{2}2-MLLM: Uncertainty-aware Revisit Reasoning in Multimodal Large Language Models for Radiology Report Generation

CliQ-RRG: Clinical-Knowledge Guided Disease-aware Visual-Textual Alignment for QA-Style Radiology Report Generation

MedScribe: Clinically Grounded CT Reporting through Agentic Workflows

UR $^{2}$ -MLLM: Uncertainty-aware Revisit Reasoning in Multimodal Large Language Models for Radiology Report Generation