The architecture of the proposed geometry-aware depth estimation system
Monocular depth estimation is critical for endoscopists to perform spatial perception and 3D navigation of surgical sites. However, most of the existing methods ignore the important geometric structural consistency, which inevitably leads to performance degradation and distortion of 3D reconstruction.A research team led by Professor LIU Hao from the Shenyang Institute of Automation (SIA) of the Chinese Academy of Science (CAS) proposed a novel geometry-aware depth estimation framework, which combined the strengths of the gradient and normal losses and geometry consistency loss. The framework enforced the geometric consistency constraints and boosted the reconstruction performance for stepped edge and frequently fluctuations structures. In addition, Liu's team introduced a synthetic RGB-D dataset tailored for interventional endoscopy. The dataset depicted the geometric structures under severe reflections and illumination variations and reduced the cost of learning surgical skills in synthetic and real-world domains.
Detailed experiments and analysis were conducted on the EndoSLAM dataset, Colondepth dataset and clinical images, indicating that the proposed method exceeded previous state-of-the-art competitors and generated more consistent depth maps and reasonable anatomical structures. The quality of intraoperative 3D structure perception from endoscopic videos of the proposed method met the accuracy requirements of video-CT registration algorithms for endoscopic navigation.
The study was published online in Engineering Applications of Artificial Intelligence (EAAI) on March 20.
