Abstract
Surgical scene understanding is constrained by data scarcity and long-tail class imbalance, as annotating surgical videos demands scarce clinical expertise while rare procedural events remain underrepresented. We tackle this through data generation and propose SurgGenesis, a generative surgical world model that learns future surgical state transitions in a latent space from visual observations and textual procedural descriptions, serving as a scalable surgical data engine. To adapt a large-scale video diffusion model to the data-scarce surgical domain without catastrophic forgetting, we introduce a three-stage progressive LoRA strategy that incrementally injects endoscopic visual priors, procedural semantics, and spatial motion cues. Rather than judging generation by visual fidelity alone, we validate SurgGenesis along two axes that separate a world model from a generic generator: (i) future-awareness, by comparing predicted futures against the real future through a GT-latent prediction gap and temporal-horizon analysis, and (ii) action-faithful usefulness, by checking that generated clips contain the commanded surgical triplet and that using them to rebalance long-tail events and drive downstream understanding (segmentation, triplet recognition, 3D reconstruction) yields action-specific gains. Conditioned futures therefore complement conventional long-tail sampling. Code: https://zh-qr.github.io/SurgGenesis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MWM_011.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=RQwPHOuMNP
BibTex
@InProceedings{ZhoQir_SurgGenesis_MICCAISAT2026,
author = { Zhong, Qirui AND Liu, Quande AND Chen, Yang AND Xue, Cheng},
title = { { SurgGenesis: A Generative Surgical World Model for Future-Aware Surgical Understanding } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17262},
month = {pending},
page = {pending}
}