Open Access

ARTICLE

Stereo-Endoscopic Disparity Estimation via Pseudo-Label-Pretrained StyleGAN3 and Self-Supervised Test-Time Optimization

Legend Zhang¹, Junmin Lyu¹, Guan Yao², Wei Wei³, Jiawei Tian^4,*, Bo Yang^2,*

1 Future Tech Institute, Guangzhou Huashang University, Guangzhou, China
2 School of Automation, University of Electronic Science and Technology of China, Chengdu, China
3 School of Mechanical and Material Engineering, Xi’an University, Xi’an, China
4 Department of Computer Science and Engineering, Hanyang University, Ansan-si, Republic of Korea

* Corresponding Authors: Jiawei Tian. Email: ; Bo Yang. Email:

(This article belongs to the Special Issue: Emerging Artificial Intelligence Technologies and Applications-II)

Computer Modeling in Engineering & Sciences 2026, 148(3), 42 https://doi.org/10.32604/cmes.2026.086568

Received 02 June 2026; Accepted 19 August 2026; Issue published 28 September 2026

Abstract

Accurate disparity estimation of soft tissue surfaces in endoscopic cardiac imaging is critical for minimally invasive surgical navigation and robotic surgery. Traditional geometric models, such as thin-plate splines, and recent learning-based stereo matching networks often struggle to balance nonlinear modeling capacity, computational cost, and robustness in dynamic surgical environments. We propose TTO-StyleGAN3 (Test-Time Optimization with a simplified StyleGAN3), a hybrid framework combining pseudo-label-supervised prior learning with self-supervised test-time latent optimization. A simplified StyleGAN3 generator is first pre-trained on pseudo-disparity maps generated by a teacher estimator and subsequently serves as a learned prior over plausible cardiac disparity fields. At inference, the generator is frozen, and the latent vector for each stereo pair is optimized without disparity supervision by minimizing photometric reconstruction, structural similarity, smoothness, and region-aware losses. This design avoids training a high-capacity end-to-end disparity predictor on manually annotated dense disparity datasets, although the compact generator prior is learned from teacher-generated pseudo-disparity maps. Experiments on a silicone heart phantom dataset and an in vivo dataset from Totally Endoscopic Coronary Artery Bypass (TECAB) procedures show that our method achieves competitive performance under the present pseudo-label and evaluation protocol, while reducing generator parameter count and the cost of training a full end-to-end disparity predictor.

---

Keywords

---