AdaEndoGS

Abstract

Endoscopic imaging technology is crucial for minimally invasive diagnosis and treatment. Due to physical constraints such as limited light source placement and narrow operational space, captured images often contain some dark regions that compromise clinical observation and diagnostic accuracy. To address this issue, we propose AdaEndoGS for endoscopic 3D reconstruction with physically consistent dark region illumination enhancement. Specifically, built upon 3D Gaussian Splatting, AdaEndoGS enriches each Gaussian point with surface attributes including normal, roughness, and reflectance, thereby constructing an illuminatable 3D scene representation. AdaEndoGS automatically identifies the dark region via ray tracing, leveraging Gaussian opacity, surface normal, and viewpoint information to adaptively plan the placement of a virtual light source. Experimental results show that AdaEndoGS more accurately simulates light-matter interactions compared to existing methods, significantly improving visual quality and enhancing detail visibility.

Method

AdaEndoGS consists of two core stages: (1) 3D Gaussian Splatting extension: enrich each Gaussian point with surface attributes (normal, roughness, reflectance) to model light-matter interaction; (2) Adaptive dark region illumination: detect dark regions via BVH-based ray tracing, then optimize virtual light source placement to enhance dark regions while maintaining physical consistency. The pipeline above shows the complete process from endoscopic image input to illumination-enhanced 3D reconstruction output.

Results & Comparison

AdaEndoGS Basic Pipeline (Input → 3D Reconstruction → Dark Region Enhancement)

Qualitative Comparison (VEIDLA Dataset, with Ground Truth)

Multi-Method Comparison (C3VD Real Endoscopic Scenes)

We evaluate AdaEndoGS on both synthetic (VEIDLA) and real-world (C3VD) endoscopic datasets, comparing with state-of-the-art methods including RetinexNet, EnlightenGAN, Aleth-NeRF and vanilla 3DGS. The qualitative results show that AdaEndoGS outperforms all baselines in dark region enhancement while preserving anatomical details and color consistency. Quantitative metrics (PSNR/SSIM/LPIPS) further validate the superiority of our method in both synthetic and real endoscopic scenes.

Ablation Study

We conduct ablation studies on the VEIDLA dataset to verify the contribution of each key component in AdaEndoGS, with evaluation metrics including PSNR (higher is better), SSIM (higher is better) and LPIPS (lower is better). The complete AdaEndoGS model achieves the optimal performance: PSNR of 22.936, SSIM of 0.908, and LPIPS of 0.165.

1) Removing DiffuseMLP (physical diffuse reflection modeling component) leads to a significant performance drop: PSNR decreases by 1.755 to 21.181, SSIM drops to 0.897, and LPIPS rises to 0.182, confirming its critical role in modeling accurate light-matter interactions for endoscopic scenes.

2) Abolishing L_tone-mse (tone mapping reconstruction loss) results in PSNR of 22.113, SSIM of 0.889 and LPIPS of 0.174. This indicates that the tone-mse loss is essential for maintaining global tonal consistency of enhanced endoscopic images.

3) Removing L_ssim (structural similarity loss) causes the most severe performance degradation: PSNR plummets to 18.371 (a 4.565 drop), SSIM reduces to 0.792, and LPIPS increases to 0.211. This fully demonstrates that the SSIM loss is the core component for preserving anatomical structural details in endoscopic dark region enhancement.

4) Abolishing L_cc (color consistency loss) leads to PSNR of 21.147, SSIM of 0.895 and LPIPS of 0.191. This verifies that the color consistency loss is key to ensuring the color fidelity of enhanced endoscopic images, avoiding color distortion in clinical scenes.