earticle

영어

The 3D reconstruction of stomach anatomy using only monocular endoscopic images presents challenges due to various factors, including illumination changes and lack of distinct surface features. Previous works have encountered difficulties in achieving dense reconstructions of stomach anatomy. In this paper, we present a patient-specific, learning-based structure- from-motion (SfM) method capable of generating dense and accurate 3D reconstructions from monocular endoscopic images. We propose a pre-processing pipeline, which is applied to an existing method for sinus anatomy reconstruction and then fine-tuned specifically for stomach anatomy using our dataset. The challenge posed by texture-less stomach surfaces is addressed by employing dense point correspondences and volumetric depth fusion. The effectiveness and accuracy of our method are demonstrated using in vivo and ex vivo data, wherein we compare sparse and dense reconstructions from the original method and our proposed method. Our qualitative results indicate that the method can accurately estimate stomach shape in real human endoscopies, thereby laying the ground for performing real-time area coverage of the stomach.