Monocular Endoscopic Tissue 3D Reconstruction with Multi-Level Geometry Regularization

Yangsen Chen; Hao Wang

cs.CVFeb 24, 2026

Monocular Endoscopic Tissue 3D Reconstruction with Multi-Level Geometry Regularization

Authors: Yangsen Chen, Hao Wang

Organizations: HKUST(GZ)

Abstract

Reconstructing deformable endoscopic tissues is crucial for achieving robot-assisted surgery. However, 3D Gaussian Splatting-based approaches encounter challenges in achieving consistent tissue surface reconstruction, while existing NeRF-based methods lack real-time rendering capabilities. In pursuit of both smooth deformable surfaces and real-time rendering, we introduce a novel approach based on 3D Gaussian Splatting. Specifically, we introduce surface-aware reconstruction, initially employing a Sign Distance Field-based method to construct a mesh, subsequently utilizing this mesh to constrain the Gaussian Splatting reconstruction process. Furthermore, to ensure the generation of physically plausible deformations, we incorporate local rigidity and global non-rigidity restrictions to guide Gaussian deformation, tailored for the highly deformable nature of soft endoscopic tissue. Based on 3D Gaussian Splatting, our proposed method delivers a fast rendering process and smooth surface appearances. Quantitative and qualitative analysis against alternative methodologies shows that our approach achieves solid reconstruction quality in both textures and geometries.

Figures & tables

Fig. 1: Our methodology begins with the key point detection, followed by neighborhood identification and mesh based Gaussian Splatting reconstruction. Subsequently, global and local restriction on the deformation is executed to obtain the final dynamic output.

Fig. 2: Comparison with other Gaussian Splatting-based methodologies. When we change the viewpoints, we clearly observe the 3D geometries of existing works are distorted, while our proposed framework reconstructs more consistent and much smoother endoscopic tissue surfaces. This demonstrates the usefulness of our proposed multi-level geometry regularization.

Fig. 3: Comparison with other methods on image quality and depth quality, where red boxes denote the incorrect reconstructed areas. Our method presents better reconstructed textures in the occluded regions.

Models	Cutting			Pulling			Training Time $\downarrow$	FPS $\uparrow$	GPU Usage $\downarrow$
Models	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Training Time $\downarrow$	FPS $\uparrow$	GPU Usage $\downarrow$
EndoNeRF	35.64	0.930	0.132	34.71	0.920	0.095	$\sim$ 6 hrs	$\sim$ 0.2	$\sim$ 20GB
EndoSurf	35.89	0.952	0.107	34.91	0.955	0.120	$\sim$ 7 hrs	$\sim$ 0.04	$\sim$ 20GB
LerPlane	34.69	0.901	0.112	36.38	0.937	0.083	$\sim$ 8 min	$\sim$ 0.9	$\sim$ 20GB
EndoGS	36.20	0.958	0.044	38.21	0.967	0.066	$\sim$ 2 min	$\sim$ 60	$\sim$ 10GB
EndoGaussian	37.21	0.961	0.065	36.10	0.946	0.091	$\sim$ 2 min	$\sim$ 170	$\sim$ 3GB

TABLE I: Comparison on the ENDONERF dataset [ 8 ] . Our approach demonstrates superior performance across most metrics and scenarios.

Models	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
EndoNeRF [ 8 ]	24.34	0.759	0.320
EndoSurf [ 22 ]	25.02	0.802	0.356
EndoGS [ 7 ]	26.47	0.798	0.291
EndoGaussian [ 26 ]	27.04	0.825	0.275
Ours	28.31	0.810	0.282

TABLE II: Comparison on the SCARED dataset [ 11 ] .

Fig. 4: Detailed comparisons of the output RGB image are presented in this section. We emphasize regions rich in blood vessels and intricate texture. Our approach demonstrates a notable capability in preserving intricate details when contrasted with alternative methodologies. Furthermore, the inpainted regions in our method exhibit a higher degree of realism and coherence.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
3DGS [ 6 ]	25.15	0.803	0.536
Dynamic 3D Gaussians [ 34 ]	33.60	0.914	0.096
Depth-Regularized GS [ 35 ]	28.38	0.913	0.472
FS-GS [ 36 ]	30.24	0.927	0.124
Ours	38.16	0.965	0.046

TABLE III: Comparison on first frame 3D reconstruction quality.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Ours	38.16	0.965	0.046
w/o surface-aware reconstruction	33.60	0.914	0.096
w/o global non-rigidity restriction	36.29	0.944	0.063
w/o local rigidity restriction	37.18	0.950	0.065

TABLE IV: Ablation study of our proposed multi-level geometry regularization on the ENDONERF dataset [ 8 ] .

Explore similar work

EndoPrior-GS: Dynamic Endoscopic Reconstruction with a Joint Texture Prior
Sep 29, 2026Jiaqi Huang, Shidong Wang, Tong Xin +13D Gaussian Splatting Medical Imaging
School of Engineering, Newcastle University, Newcastle upon Tyne NE1 7RU, UK · School of Computing, Newcastle University, Newcastle upon Tyne NE4 5TG, UK
Diff2DGS: Reliable Reconstruction of Occluded Surgical Scenes via 2D Gaussian Splatting
Feb 20, 2026Tianyi Song, Danail Stoyanov, Evangelos Mazomenos +13D Reconstruction Dynamic Scene Reconstruction
EndoVGGT: GNN-Enhanced Depth Estimation for Surgical 3D Reconstruction
Mar 25, 2026Falong Fan, Yi Xie, Arnis Lektauers +23D Reconstruction Depth Estimation
University of Arizona, Tucson, AZ, USA

Sep 29, 2026J/K move · Enter open · S save

EndoPrior-GS: Dynamic Endoscopic Reconstruction with a Joint Texture Prior

Jiaqi Huang, Shidong Wang, Tong Xin, Kabita Adhikari

School of Engineering, Newcastle University, Newcastle upon Tyne NE1 7RU, UK · School of Computing, Newcastle University, Newcastle upon Tyne NE4 5TG, UK

3D Gaussian Splatting Medical Imaging Dynamic Scene Reconstruction Deformable Shape Reconstruction

Open paper AskPDF arXiv

Dynamic endoscopic reconstruction is fundamental to robotic surgery and computer-assisted interventions. While 3D Gaussian Splatting (3DGS) realises real-time rendering, its application to deformable intraoperative environments remains constrained by spurious geometry and varying illuminations. To address these limitations, we introduce EndoPrior-GS, a novel pipeline that explicitly couples frame-extracted vision heuristics and estimated depth maps. EndoPrior-GS derives a joint texture prior from a tool-filtered valid tissue mask, a non-specular photometric filter, and anatomical structural salience, yielding a probability map that guides primitive initialisation and subsequent density control. The prior is further extended to the temporal domain through a texture-aware term that dynamically weighs pairwise primitive contributions during training. We conduct extensive experiments on benchmark datasets EndoNeRF and SCARED, and the obtained results show that our method EndoPrior-GS reduces Flow Error by 27.7% and 25.8% over the representative approaches while preserving competitive rendering quality and real-time rendering speed. Our project website is available at https://jiaqi-huang-77.github.io/EndoPrior-GS/.

Abstract

Explore similar work

EndoPrior-GS: Dynamic Endoscopic Reconstruction with a Joint Texture Prior

Diff2DGS: Reliable Reconstruction of Occluded Surgical Scenes via 2D Gaussian Splatting

EndoVGGT: GNN-Enhanced Depth Estimation for Surgical 3D Reconstruction