Source-linked AI summary
Structured Reasoning Agentic Framework for Interpretable Critical View of Safety Assessment
Qing Xu, Yuxiang Luo, Zhen Chen
TL;DR
CVS assessment methods often map visual features directly to criterion-level labels, without explicit reasoning about anatomical relationships. ReasonCVS structures assessment as graph-based anatomical verification with sub-criterion checks and rationale-aware synthesis. On Endoscapes-CVS201, it achieves superior performance over state-of-the-art methods while producing traceable clinical rationales.
Problem
Existing CVS methods typically make holistic criterion-level predictions without explicit anatomical-relation reasoning, limiting interpretability and compositional generalization.
Source: §1 · p. 2Method
ReasonCVS represents anatomical entities and spatial relationships as a graph, verifies individual sub-criteria with VLMs, and uses a rationale-distilled agent to synthesize observations.
Source: §1 · p. 2, §2.1 · p. 4Results
Takeaways & Limitations
The framework provides explicit anatomical verification and criterion-level explanations with traceable clinical rationales for CVS assessment.
Source: §1 · pp. 2–3
Abstract
from arXiv · showhide
Surgical scene understanding is critical for computer-assisted intervention, yet laparoscopic cholecystectomy remains challenged by the complex anatomy of the hepatocystic triangle and the risk of bile duct injury. Existing methods for Critical View of Safety (CVS) assessment typically treat it as a holistic prediction task, mapping visual features directly to criterion-level labels. This black-box paradigm lacks explicit reasoning about anatomical relationships, limiting both interpretability and compositional generalization. To address this, we propose ReasonCVS, a structured reasoning agentic framework empowered by Vision-Language Models (VLMs) that decomposes CVS assessment into explicit, fine-grained anatomical verification. Specifically, we devise an Anatomical Scene Graph Abstraction (ASGA) that organizes anatomical entities and their spatial relationships into a structured representation. To operationalize this, we introduce a Rationale-Aware Reasoning Agent, powered by a Large Language Model (LLM) fine-tuned via rationale distillation. Functioning as a strict central decision-maker, it invokes VLM-driven Sub-criterion Verifier as a specialized perceptual tool to parse the graph and independently evaluate individual sub-criteria. Through calibrated soft reasoning, this agent synthesizes the tool-gathered distributed observations, yielding a final verdict alongside a traceable clinical rationale. Extensive experiments on the Endoscapes-CVS201 benchmark demonstrate that ReasonCVS achieves superior performance (68.1\% mAP) over state-of-the-art while providing interpretable, criterion-level explanations for reliable surgical assessment.
1 Introduction
Laparoscopic cholecystectomy requires careful identification of hepatocystic-triangle structures to prevent bile duct injury, but existing CVS methods often predict criterion labels holistically without explicit anatomical reasoning. ReasonCVS instead decomposes assessment into fine-grained verification of anatomical relationships, with a traceable clinical rationale.
1 Introduction
- Laparoscopic cholecystectomy is challenged by complex hepatocystic-triangle anatomy and the need to identify critical structures to prevent bile duct injury.These challenges motivate automated Critical View of Safety assessment.Source: §1 · p. 2
- Holistic CVS prediction maps visual features directly to criterion labels, but lacks explicit anatomical-relation reasoning, limiting interpretability and compositional generalization.CVS assessment depends on relationships among hepatocystic-triangle structures, such as verifying that only the cystic duct and cystic artery enter the gallbladder.Source: §1 · p. 2
- ReasonCVS decomposes CVS assessment into explicit, fine-grained anatomical verification using a structured scene graph and sub-criterion checks.The framework overview also includes a Rationale-Aware Reasoning Agent to produce a traceable clinical rationale.Source: §1 · p. 2, Figure 2 · p. 3
- A Rationale-Aware Reasoning Agent, fine-tuned via rationale distillation, centrally synthesizes verifier observations through calibrated soft reasoning into a verdict and traceable clinical rationale.The authors report that experiments demonstrate superior performance over state-of-the-art methods.Source: §1 · pp. 2–3
2 Methodology
ReasonCVS recasts CVS assessment as explicit verification of anatomical relationships, then synthesizes VLM-based sub-criterion observations into traceable criterion-level decisions. Its methodology combines an anatomical scene graph, specialized verifiers, and a rationale-aware agent.
2 Methodology
- ReasonCVS combines graph-based anatomical representation, VLM-driven sub-criterion verification, and an agent that synthesizes distributed evidence through calibrated soft reasoning.The framework decomposes CVS assessment into explicit verification of fine-grained anatomical relationships.Source: §2 · p. 3
2.1 Anatomical Scene Graph Abstraction
- ASGA defines a reference graph whose nodes represent expected structures and whose edges encode required spatial relationships and fine-grained constraints.It builds the observed graph from segmentation masks, then enriches it with a colored overlay and local instance crops.Source: §2.1 · p. 4
2.2 VLM-driven Sub-criterion Verification
- VLM verifiers parse criteria into node-attribute or edge-property checks and return judgments, confidence scores, and evidence-grounded explanations.Each verifier receives the annotated overlay, relevant local crops, and structured graph; the resulting observations are forwarded for criterion-level synthesis.Source: §2.2 · p. 4
2.3 Rationale-Aware Reasoning Agent
- The rationale-aware agent synthesizes verifier assessments into criterion verdicts and traceable rationales, upweighting high-confidence evidence and resolving conflicts.Its rationale references specific sub-criterion evidence so decisions can be traced to inspectable observations.Source: §2.3 · p. 5
2.4 Optimization Pipeline
- Verifiers are prompted without additional training, while the reasoning agent is fine-tuned through teacher-student rationale distillation.Because the dataset lacks sub-criterion ground truth, the teacher generates evidence-to-verdict rationales and the student learns from those triples.Source: §2.4 · p. 5
3 Experiments
Experiments compare ReasonCVS with prior CVS assessment methods, test its components through ablations, and evaluate its reasoning quality with expert scores.
3.1 Experimental Setup
- Experiments use Endoscapes-CVS201, with 11,090 frames annotated for image-level CVS criteria and 493 frames with pixel-level segmentation masks.The evaluation uses mAP and Balanced Accuracy (BAcc), and follows the official train, validation, and test splits.Source: §3.1 · p. 6
3.2 Comparison with State-of-the-art Methods
- 68.1% mAP makes ReasonCVS the best-performing method, surpassing SV2LSTG by 3.8% and reaching 81.4% BAcc (+6.6% over LG-CVS).On C2, ReasonCVS scores 62.9% versus 56.4%.Source: §3.2 · p. 7
3.3 Ablation Study
- Adding ASGA improves performance by 10.6% mAP and 8.3% BAcc; adding VSV yields further gains of 6.1% mAP and 7.6% BAcc.The authors connect VSV’s gains to more precise anatomical assessment through fine-grained sub-criterion decomposition.Source: §3.3 · p. 7
- Integrating RARA brings the largest component gains: 7.8% mAP and 9.0% BAcc.The authors attribute these gains to rationale distillation teaching the agent to weigh uncertain evidence and resolve conflicts across sub-criteria.Source: §3.3 · p. 8
3.4 Clinical Evaluation of Reasoning Quality
- ReasonCVS receives higher and more concentrated expert scores than the Vanilla Agent, with the clearest separation on Clinical Reasonableness (4–5 versus 1–2).A certified surgeon scored 30 samples, 10 per criterion, on a 1–5 Likert scale.Source: §3.4 · p. 8, Figure 4 · p. 8
4 Conclusion
ReasonCVS combines structured anatomical representations, sub-criterion verification, and rationale-aware agentic reasoning for CVS assessment. On Endoscapes-CVS201, it achieves 68.1% mAP while providing criterion-level explanations.
4 Conclusion
- ReasonCVS organizes anatomical entities and their spatial relationships in an Anatomical Scene Graph Abstraction, then verifies individual sub-criteria with a VLM-driven verifier.A rationale-distilled Rationale-Aware Reasoning Agent synthesizes the observations through calibrated soft reasoning into a verdict and traceable clinical rationale.Source: §4 · p. 8
- ReasonCVS achieves 68.1% mAP on Endoscapes-CVS201, outperforming state-of-the-art methods.The framework provides interpretable, criterion-level explanations.Source: §4 · p. 8
- The conclusion presents ReasonCVS as an approach to reliable surgical assessment that pairs performance with interpretable explanations.Source: §4 · p. 8