Dhakal, Rochak, Zhao, Chen, Cobb, Soup, Qian, Xin, Yang, Xiaofeng, Zhou, Weihua · Nov 6, 2026
GRAC: Graph-Based Retrieval-Augmented Captioning for Surgical Scene Description
Markdown · JSON · Original source
Rochak Dhakal, Chen Zhao, Soup Cobb, Xin Qian, Xiaofeng Yang and Weihua Zhou propose GRAC, a system for generating descriptions of surgical scenes. It represents instruments, actions and anatomical structures as graphs, then retrieves related graph-caption pairs from training data. The system compares instrument-tissue interactions in retrieved examples with those in the input scene and uses those comparisons when generating a caption. The authors say automated descriptions could reduce the manual review and clinical expertise needed to extract information for operative reports, training, performance assessment and workflow review. On the EndoVis-18 , the authors report higher CIDEr-D and exact instrument-action-target triplet-set accuracy for GRAC than for the named comparison systems. On TORS, they also report higher CIDEr-D than the strongest baseline. The authors conclude that combining structured surgical scene representations with retrieved examples improves surgical scene captioning and produces more contextually meaningful descriptions. These are benchmark results, not evidence here of performance in clinical use. Whether that approach remains effective on surgical scenes beyond the reported benchmarks is unresolved.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 6, 2026 (2026-11-06).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T04:17:11.549Z. First observed: 2026-11-09T05:06:56.811Z. Retrieved: 2026-11-10T04:15:56.389Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the paper’s abstract, which reports results but does not describe performance in clinical use. The abstract does not provide details of how GRAC performs on surgical scenes beyond the EndoVis-18 and TORS benchmarks.
Open question
How does GRAC perform on surgical scenes beyond the EndoVis-18 and TORS reported by the authors?
Source-stated
The authors propose GRAC, which represents surgical scenes as graphs and retrieves related graph-caption pairs to inform caption generation.
Given an input scene graph, GRAC learns to retrieve relevant graph-caption pairs from training data.
Source-stated
On EndoVis-18, the authors report GRAC CIDEr-D of 5.88, versus 4.23 for SGT and 3.36 for the strongest conventional Transformer baseline.
On the EndoVis-18 benchmark, GRAC achieved a CIDEr-D of 5.88, compared with 4.23 for SGT and 3.36 for the strongest conventional Transformer baseline.
Source-stated
On EndoVis-18, the authors report exact instrument-action-target triplet-set accuracy of 52.6% for GRAC, versus 32.7% for SGT and 26.2% for the strongest conventional baseline.
Exact instrument-action-target triplet-set accuracy reached 52.6%, compared with 32.7% for SGT and 26.2% for the strongest conventional baseline.
Source-stated
On TORS, the authors report GRAC CIDEr-D of 5.06, compared with 3.80 for the strongest baseline.
On TORS, GRAC achieved a CIDEr-D of 5.06, compared with 3.80 for the strongest baseline.
Source-stated
The authors say automated surgical scene descriptions could reduce the burden of extracting information for documentation, training, performance assessment and workflow review; the abstract does not establish that reduction in practice.
Automated surgical scene descriptions could reduce this burden by leveraging recurring surgical activities and prior knowledge.
Source and contributor context
Dhakal, Rochak, Zhao, Chen, Cobb, Soup, Qian, Xin, Yang, Xiaofeng, Zhou, Weihua · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 4de1fa33-f060-418f-8fe0-6f96fd840579.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T04:17:11.549Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: bd019b43934e39aea27403e3590599f0842ead07c7f5d5458f20d960980e8563. Source text hash: fa1c8d52266ac4f15ca8826520988c8b06490c50190e0167b9f4fce63413c272.
Revision and source-check history
Version 1 · 2026-11-10T04:17:11.549Z
GRAC: Graph-Based Retrieval-Augmented Captioning for Surgical Scene Description
Rochak Dhakal, Chen Zhao, Soup Cobb, Xin Qian, Xiaofeng Yang and Weihua Zhou propose GRAC, a system for generating descriptions of surgical scenes. It represents instruments, actions and anatomical structures as graphs, then retrieves related graph-caption pairs from training data. The system compares instrument-tissue interactions in retrieved examples with those in the input scene and uses those comparisons when generating a caption. The authors say automated descriptions could reduce the manual review and clinical expertise needed to extract information for operative reports, training, performance assessment and workflow review. On the EndoVis-18 , the authors report higher CIDEr-D and exact instrument-action-target triplet-set accuracy for GRAC than for the named comparison systems. On TORS, they also report higher CIDEr-D than the strongest baseline. The authors conclude that combining structured surgical scene representations with retrieved examples improves surgical scene captioning and produces more contextually meaningful descriptions. These are benchmark results, not evidence here of performance in clinical use. Whether that approach remains effective on surgical scenes beyond the reported benchmarks is unresolved.
Original source
cs.AI · Article
Source details
Original source date: Nov 6, 2026 (2026-11-06).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T04:17:11.549Z. First observed: 2026-11-09T05:06:56.811Z. Retrieved: 2026-11-10T04:15:56.389Z.
Limits
The supplied evidence is the paper’s abstract, which reports results but does not describe performance in clinical use. The abstract does not provide details of how GRAC performs on surgical scenes beyond the EndoVis-18 and TORS benchmarks.
Open question
How does GRAC perform on surgical scenes beyond the EndoVis-18 and TORS reported by the authors?
Source-stated
The authors propose GRAC, which represents surgical scenes as graphs and retrieves related graph-caption pairs to inform caption generation.
Given an input scene graph, GRAC learns to retrieve relevant graph-caption pairs from training data.
Source-stated
On EndoVis-18, the authors report GRAC CIDEr-D of 5.88, versus 4.23 for SGT and 3.36 for the strongest conventional Transformer baseline.
On the EndoVis-18 benchmark, GRAC achieved a CIDEr-D of 5.88, compared with 4.23 for SGT and 3.36 for the strongest conventional Transformer baseline.
Source-stated
On EndoVis-18, the authors report exact instrument-action-target triplet-set accuracy of 52.6% for GRAC, versus 32.7% for SGT and 26.2% for the strongest conventional baseline.
Exact instrument-action-target triplet-set accuracy reached 52.6%, compared with 32.7% for SGT and 26.2% for the strongest conventional baseline.
Source-stated
On TORS, the authors report GRAC CIDEr-D of 5.06, compared with 3.80 for the strongest baseline.
On TORS, GRAC achieved a CIDEr-D of 5.06, compared with 3.80 for the strongest baseline.
Source-stated
The authors say automated surgical scene descriptions could reduce the burden of extracting information for documentation, training, performance assessment and workflow review; the abstract does not establish that reduction in practice.
Automated surgical scene descriptions could reduce this burden by leveraging recurring surgical activities and prior knowledge.
Source and contributor context
Dhakal, Rochak, Zhao, Chen, Cobb, Soup, Qian, Xin, Yang, Xiaofeng, Zhou, Weihua · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 4de1fa33-f060-418f-8fe0-6f96fd840579.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T04:17:11.549Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: bd019b43934e39aea27403e3590599f0842ead07c7f5d5458f20d960980e8563. Source text hash: fa1c8d52266ac4f15ca8826520988c8b06490c50190e0167b9f4fce63413c272.
passed · 2026-11-10T04:17:11.549Z · anthropic/claude-opus-5.5
Publication: 4de1fa33-f060-418f-8fe0-6f96fd840579
attribution — pass: The reading names the six listed authors correctly and attributes the proposal, the benchmark results and the conclusion to them. The potential burden reduction is framed as the authors' view, not as an established fact.
Authors:Rochak Dhakal, Chen Zhao, Soup Cobb, Xin Qian, Xiaofeng Yang, Weihua Zhou
summary — pass: The summary faithfully describes the method, motivation, results and conclusion, and it notes that these are benchmark results only. The question asks about one real gap. Small repair: the abstract says the graphs capture relationships among instruments, actions and anatomy; restoring that word would make 'graphs' clearer. CIDEr-D, EndoVis-18, TORS and triplet-set accuracy would benefit from term-library tooltips. They are not explained in the source, but the reader can still see that GRAC scored higher on the reported tests.
represents surgical scenes as graphs capturing relationships among instruments, actions, and anatomical structures
date — pass: originalPublished is 2026-11-06, which matches the arXiv v1 submission date. The feed date of Nov 9 is correctly not used.
[Submitted on 6 Nov 2026]
support — pass: All claims match the abstract's figures exactly and are attributed to the source. The reading explicitly says the benefit is not established in practice and does not claim clinical reliability. Its unresolved question follows from the evidence and does not invent stakeholder impacts.
On TORS, GRAC achieved a CIDEr-D of 5.06, compared with 3.80 for the strongest baseline.
media — pass: The selected media is the verified direct link to the original arXiv abstract page, which is the source itself. No better media was found.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference