[Paper Review] EndoGS: Deformable Endoscopic Tissues Reconstruction with Gaussian Splatting

Lingting Zhu, Zhao Wang+4 more|arXiv (Cornell University)|Jan 21, 2024

3D Shape Modeling and Analysis4 citations

TL;DR

EndoGS proposes the first Gaussian Splatting-based method for high-fidelity, real-time reconstruction of deformable endoscopic tissues from single-view videos. By integrating deformation fields, depth-guided supervision with spatial-temporal masks, and surface-aligned regularization, it achieves state-of-the-art rendering quality and ~70 FPS, significantly outperforming prior NeRF-based methods in both quality and speed.

ABSTRACT

Surgical 3D reconstruction is a critical area of research in robotic surgery, with recent works adopting variants of dynamic radiance fields to achieve success in 3D reconstruction of deformable tissues from single-viewpoint videos. However, these methods often suffer from time-consuming optimization or inferior quality, limiting their adoption in downstream tasks. Inspired by 3D Gaussian Splatting, a recent trending 3D representation, we present EndoGS, applying Gaussian Splatting for deformable endoscopic tissue reconstruction. Specifically, our approach incorporates deformation fields to handle dynamic scenes, depth-guided supervision with spatial-temporal weight masks to optimize 3D targets with tool occlusion from a single viewpoint, and surface-aligned regularization terms to capture the much better geometry. As a result, EndoGS reconstructs and renders high-quality deformable endoscopic tissues from a single-viewpoint video, estimated depth maps, and labeled tool masks. Experiments on DaVinci robotic surgery videos demonstrate that EndoGS achieves superior rendering quality. Code is available at https://github.com/HKU-MedAI/EndoGS.

Motivation & Objective

• To address the challenge of reconstructing high-quality, dynamic endoscopic tissues from single-viewpoint videos, which is ill-posed due to occlusions and large deformations.
• To overcome limitations of existing NeRF-based methods, such as slow optimization and poor generalization under tool occlusion.
• To leverage the speed and fidelity of 3D Gaussian Splatting for surgical 3D reconstruction, enabling real-time applications.
• To improve geometric accuracy in reconstructed scenes, especially near tool-occluded regions, through surface-aligned regularization.

Proposed method

• Represents dynamic surgical scenes using a combination of static 3D Gaussians and a time-dependent deformation field modeled via a lightweight MLP.
• Applies depth-guided supervision with spatial-temporal weight masks to handle tool occlusion during monocular optimization.
• Introduces surface-aligned regularization to enhance geometric fidelity by aligning Gaussians with underlying tissue surfaces.
• Uses differentiable rasterization to render novel views and optimize against ground-truth images, depth maps, and tool masks.
• Employs a two-stage training strategy: first optimizing static Gaussians, then refining the deformation field with additional iterations.
• Leverages multi-resolution voxel planes to extract features from Gaussians' mean and time for deformation prediction.

Experimental results

Research questions

• RQ1 Can Gaussian Splatting be effectively adapted to reconstruct deformable endoscopic tissues from single-view videos with high fidelity and real-time performance?
• RQ2 How can tool occlusion in monocular endoscopic videos be mitigated during 3D reconstruction using implicit or explicit supervision?
• RQ3 What role does surface-aligned regularization play in improving geometric accuracy and reducing artifacts near occluded regions?
• RQ4 How does depth-guided supervision with spatial-temporal masks improve reconstruction quality under partial visibility?
• RQ5 Does combining deformation modeling with 3D Gaussian Splatting outperform existing NeRF-based approaches in both rendering quality and inference speed?

Key findings

• EndoGS achieves a PSNR of 37.935 (±0.088), SSIM of 0.966 (±0.003), and LPIPS of 0.034 (±0.001), significantly outperforming EndoNeRF and ForPlane in all metrics.
• The method achieves approximately 70 FPS, a substantial improvement over EndoNeRF (<0.2 FPS) and ForPlane (~1.7 FPS), demonstrating real-time capability.
• Abllation studies confirm that spatial Total Variation loss reduces color drift in occluded regions, improving visual consistency.
• Surface-aligned regularization effectively reduces blurring and preserves fine details near tool-occluded surfaces, as shown in qualitative comparisons.
• The use of consistent tool masks in training and inference ensures fair comparison and enhances robustness to occlusion.
• EndoGS demonstrates superior generalization on challenging scenes with large non-rigid deformations, such as 'traction' and 'pulling soft tissues'.