Abstract

Instrument–action–target triplets compactly describe interactions between surgical tools and anatomy. Vision–language models (VLMs) can reason over these relations but may exhaust generation budgets before producing a valid decision. We introduce SurgTriKG, which compiles image-conditioned event proposals weakly anchored by available SurgΣ annotations into a canonical event knowledge graph (KG). Query-conditioned statistics guide VLM reasoning without parameter updates. Across 40,525 de-smoked frames, the compilation audit yields 97,036 canonical relation occurrences, while all downstream graph context is derived exclusively from the training videos. On held-out videos, graph context raises Triplet Selection accuracy from 19.1% to 47.8% and Macro-F1 from 27.7% to 52.5% at a 2,048-token limit. It also reduces parse failures from 62.4% to 18.7%. Triplet Verification Macro-F1 improves by 4.2–7.9 points across four budgets, with gains concentrated in positive-class F1 and recall, indicating enhanced support sensitivity. On video-disjoint splits, a diagnostic probe reaches 98.5% Macro-F1. KG features alone reach 97.8% and retain 95.7% after removing explicit entity-hit indicators. These results support canonical event graphs as structured context for surgical triplet reasoning.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EndoLINA_008.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=NQhuvHLwfB

BibTex

@InProceedings{DonJin_SurgTriKG_MICCAISAT2026,
        author = { Dong, Jingqi AND Hu, Yudong AND Fu, Suzhong AND Bao, Lunhong AND Ding, Xuan AND Sun, Rui AND Yang, Yizhou AND Cui, Shuguang AND Li, Zhen},
        title = { { SurgTriKG: Canonical Event Graphs for Training-Free Surgical Triplet Reasoning } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17275},
        month = {pending},
        page = {pending}
}


back to top