EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

OcularWorldSim: A Visual World Model for Ophthalmic Microsurgery With Sim-to-Real Transfer Learning

Hongrui Chen, Lijuan Zhu, Lihong Lian, Yuqi Bai, Yuzhou Liu

IEEE Access · 2026 · 14 · 99426-99439

DOI: 10.1109/access.2026.3684562 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Hongrui Chen und 4 weiteren Autor:innen. Die Arbeit nennt 62 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Microsurgery involving the eye requires precise surgical techniques performed at micrometer accuracy, making it difficult to design a robotic or autonomous surgical system that will work effectively in the domain of ophthalmic microsurgery. Some advances have been made toward surgical robotics and the field of Artificial Intelligence has seen great success, however, the methods used in solving the "sim-to-real" problem, which deals with bridging the gap between the simulated training environment and clinical environment, have not made sufficient progress. In this paper, we present OcularWorldSim, a new visual world model framework for ophthalmic microsurgery that provides a means for effective "sim-toreal" learning using simulations and clinical data. The OcularWorldSim framework consists of three core innovations: (1) an Anatomically-Aware Spatiotemporal Diffusion Transformer (AA-SDiT), which learns a structurally organized representation of surgical dynamics while also maintaining the constraints of eye anatomy; (2) a Hierarchical Domain Adaptation Module (HDAM) which progressively aligns the simulated and actual surgical distributions at different levels of granularity; and (3) a Clinically-Guided Reinforcement Learning (CGRL) paradigm that incorporates expert surgical knowledge into the optimization process used for developing the policy. In the area of surgical phase recognition using CATARACTS, OcularWorldSim achieved an accuracy of 89.7% (5.8 percentage points higher than the previously best performing method) while achieving 82.4% mean Intersection-over-Union (mIoU) across CaDIS tasks (7.4 percentage points higher). Clinical validation of the predictions produced by the OcularWorldSim framework performed by three experienced ophthalmologists showed that all predictions made by the OcularWorldSim framework are temporally coherent and anatomically plausible, indicating a potential future application for the OcularWorldSim framework in computer-assisted intervention systems.

Förderung

  • National Natural Science Foundation of China

Zitieren

APA

Chen, H., Zhu, L., Lian, L., Bai, Y., & Liu, Y. (2026). OcularWorldSim: A Visual World Model for Ophthalmic Microsurgery With Sim-to-Real Transfer Learning. IEEE Access, 14, 99426-99439. https://doi.org/10.1109/access.2026.3684562

Vancouver

Chen H, Zhu L, Lian L, Bai Y, Liu Y. OcularWorldSim: A Visual World Model for Ophthalmic Microsurgery With Sim-to-Real Transfer Learning. IEEE Access. 2026;14:99426-99439. doi:10.1109/access.2026.3684562

BibTeX
@article{Chen2026,
  author = {Chen, Hongrui and Zhu, Lijuan and Lian, Lihong and Bai, Yuqi and Liu, Yuzhou},
  title = {OcularWorldSim: A Visual World Model for Ophthalmic Microsurgery With Sim-to-Real Transfer Learning},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {99426--99439},
  doi = {10.1109/access.2026.3684562}
}
Kostenlos starten →