Core Concepts
Utilizing near-field lighting for improved monocular depth estimation in endoscopy videos.
Abstract
The content discusses the challenges in depth estimation in endoscopy videos due to illumination effects and lack of geometric features. It introduces a novel approach using per-pixel shading representation and a depth refinement network (PPSNet) to enhance depth estimation. The teacher-student transfer learning method is proposed to improve depth maps from synthetic and clinical data. State-of-the-art results are achieved on the C3VD dataset, showcasing the effectiveness of the approach.
Stats
"In 2022 an estimated 225.3 million endoscopic procedures were performed globally with an annual growth rate of 1.3%."
"Our method outperforms the state-of-the-art monocular depth estimation technique LightDepth [54]."
"Our method is trained with only ∼18,000 clinical endoscopy images using lighting-based self-supervision."
Quotes
"Our key idea is to model the reflection of co-located light with the camera of the endoscope from the surface of the internal organ to help predict the relative distance between the endoscope and the surface."
"We show that our proposed approach significantly outperforms state-of-the-art monocular depth estimation techniques on the synthetic C3VD dataset."
"Our method produces higher-quality depth maps with fewer discrepancies in the case of C3VD and clinical data."