Fast and Robust 3D Reconstruction of Anatomical Cavities

Abstract

Neural Radiance Fields (NeRF) offers great potential for 3D medical reconstruction, but faces challenges in monocular endoscopy due to motion blur and non-uniform lighting, leading to significant artifacts in the final surface. We present an adapted pipeline using Instant-NGP that addresses these issues by decoupling pre-processing from neural training. Our workflow incorporates geometric calibration for rectification, targeted de-blurring and robust pose estimation to prepare frames for rapid training. We evaluate our Novel View Synthesis (NVS) on custom lung and knee models, demonstrating visually coherent novel view synthesis in under 5 minutes with high image quality and structural accuracy.