CrossScope: A Role-Asymmetric World Model for Joint Dual-Scope Surgical Video Prediction

Problem setting

Abstract

Visual world models typically learn future dynamics from a single observation stream, limiting their ability to model cooperative systems with multiple independently moving observers. We investigate this challenge in Mother-Child endoscopic retrograde cholangiopancreatography (ERCP), where two flexible scopes provide complementary yet role-dependent views without a calibrated stereo relationship.

We formulate Role-asymmetric dual-scope future prediction, where cross-view evidence is selectively transferred according to the prediction target and its spatial requirements. CrossScope is a dual-stream surgical world model that preserves view-specific experts while enabling target-specific evidence routing through geometry-guided residual interactions.

Geometric motion cues from the Mother view guide Child-view future dynamics, while pose-aligned Child appearance supports Mother-view prediction only when valid spatial correspondence is established. On paired phantom and real-world ERCP episodes, CrossScope improves visual fidelity, structural preservation, target localization, and motion consistency over surgical video-generation baselines.

Role-asymmetric routing

Why role asymmetry?

The Mother scope supplies the stable global map, while the Child scope brings close-range detail from narrow bile ducts. Their fields of view, scale, and pose evolve independently during ERCP.

CrossScope turns this mismatch into two directed routes: M2C sends pose-derived motion evidence to the Child expert, and C2M injects Child appearance only where Mother-plane geometry supports the correspondence.

2 independently controlled scopes M2C geometry-guided motion C2M pose-aligned appearance

CrossScope qualitative prediction comparison across Mother and Child views.

Pose signal Pose signal used to align the dual-scope routing.

Motion route Motion experiment visualization for Mother-to-Child routing.

C2M support Spatial support diagnostic for Child-to-Mother appearance routing.

Model design

CrossScope architecture

Separate DiT experts read a shared pre-injection snapshot and receive only target-valid residual writes.

CrossScope dual-stream architecture with M2C geometric motion conditioning and C2M pose-aligned appearance conditioning. — **Target-specific residual exchange.** Pose Readout maps Mother states to a Mother-plane trajectory for M2C. C2M pose-aligns Current and causal History Child observations before writing or blending Mother residuals under geometric support.

Quantitative evidence

End-to-end results on the phantom benchmark

Frame fidelity across Child and Mother views. Higher PSNR/SSIM and lower FID/LPIPS are better.

Model	Child PSNR	Child SSIM	Child FID	Child LPIPS	Mother PSNR	Mother SSIM	Mother FID	Mother LPIPS
HunyuanVideo-I2V	34.332	0.942	18.213	0.0848	35.112	0.953	34.324	0.0352
Cosmos-H-Surgical	35.123	0.943	17.431	0.0821	35.774	0.969	33.617	0.0347
Wan2.2	35.003	0.940	17.986	0.0854	35.001	0.956	34.618	0.0355
Early-Fusion Wan2.2	30.790	0.915	20.190	0.0877	35.432	0.966	33.378	0.0354
Symmetric Cross-Attention MoT	33.326	0.931	18.369	0.0852	36.298	0.966	29.019	0.0351
CrossScope	35.906	0.951	17.402	0.0819	37.147	0.973	28.232	0.0334

Child-scope motion quality

Papilla trajectory error on the phantom benchmark. Lower is better; values are in pixels.

Model	Endpoint error (px) ↓	Centerline error (px) ↓
HunyuanVideo-I2V	36.068	19.126
Cosmos-H-Surgical	28.880	15.339
Wan2.2	35.093	18.544
Early-Fusion Wan2.2	42.590	21.733
Symmetric Cross-Attention MoT	31.767	12.414
CrossScope	19.343	11.238

CrossScope also achieves 0.948 Mother mask IoU, 0.839 Child papilla box IoU, and 0.927 detection recall.

Routing validity

Geometry-licensed C2M support

C2M has a different information contract from M2C: it writes only pose-aligned appearance from observed Child frames to the Mother expert. Current is the observed Child anchor; History contains strictly earlier cached observations.

Support is established before a residual is written. The predicted trajectory can select a Current-only, History-only, dual-source, or Null state without admitting future Child-token states as appearance evidence.

Four C2M spatial support cases showing the Mother anchor, Current coverage, causal History coverage, and final injection state. — Four C2M spatial-support cases: the routing state changes with the geometrically supported Current and causal History coverage.

Data coverage

Dual-scope observations across phantom and real-world settings

Representative Mother, mask, and Child observations make the coupled viewpoint relationship visible in both the phantom benchmark and real-world procedures.

Representative Mother, mask, and Child observations across phantom and real-world settings. — Paired dual-scope observations across phantom and real-world settings.

Reference

Citation

@misc{crossscope2026,
  title={CrossScope: A Role-Asymmetric World Model for Joint Dual-Scope Surgical Video Prediction},
  author={Wanhao Liu and Jinsong Lin and Rulin Zhou and CHI KIT NG and Wenbin Pan and Zhiqing Tang and Dongyue Li and Miao Luo and Wu Yanshen and Panshuo Li and Zhiyong Xiong and Huxin Gao and Tamas Haidegger and Hongliang Ren},
  year={2026}
}