AdaEndoGS: An Adaptive Enlightening Model for Endoscopy Based on 3D Gaussian Splatting

Area of Science:

• Medical Imaging
• Computer Vision
• 3D Reconstruction

Background:

• Endoscopic imaging is vital for minimally invasive procedures.
• Dark regions in endoscopic scenes hinder accurate diagnosis.
• Current reconstruction methods struggle with poor illumination.

Purpose of the Study:

• To develop an advanced endoscopic 3D reconstruction technique.
• To enhance dark regions in endoscopic images with physically consistent illumination.
• To improve diagnostic accuracy through better visual quality.

Main Methods:

• Proposed AdaEndoGS, building upon 3D Gaussian Splatting.
• Enriched Gaussian primitives with surface attributes (normal, roughness, reflectance).
• Implemented adaptive virtual light source placement and optimized illumination parameters via a novel loss function.

Main Results:

• AdaEndoGS accurately simulates light-matter interactions.
• Significantly improved visual quality and detail visibility in dark endoscopic regions.
• Outperformed existing methods in quantitative and qualitative evaluations.

Conclusions:

• AdaEndoGS offers an effective solution for endoscopic 3D reconstruction and rendering.
• Physically consistent dark region enhancement improves clinical observation.
• The method advances endoscopic image-based diagnosis and treatment.