単眼大腸内視鏡検査におけるカメラ内因性ガイド学習によるスケール整合性のある3D再構成

Reliable spatial understanding of the colon lumen is crucial for geometry-aware analysis and quality assessment in colonoscopy. However, monocular colonoscopes are not equipped with depth sensors, and existing methods struggle to recover depth and camera motion with a stable, consistent scale. To address this challenge, we propose a novel framework for 3D reconstruction in monocular colonoscopy, aiming to recover scale-consistent depth and camera motion in an intrinsics-defined scale space. The framework adopts a two-step training strategy. Step 1, a depth estimation network is pretrained on synthetic colonoscopy data derived from CT-based colon models. We introduce an Adaptive Depth Adjustment module to better match endoscopic depth characteristics and to produce more stable depth predictions. The pretrained model is then applied to real images to predict depth under the synthetic-camera setting, and an intrinsics-guided scale alignment module converts these predictions into pseudo-depth labels that are interpretable in the target intrinsics-defined scale space. Step 2, these pseudo-labels guide joint depth-pose learning on real sequences with a geometry-aware 3D loss to enforce multi-view consistency and suppress scale drift. During inference, a globally consistent optimization strategy mitigates long-sequence drift. Experiments on C3VD show state-of-the-art depth and pose accuracy, reducing AbsRel by 16.1% and trajectory APE by 86.7% over the strongest baseline. Evaluations on clinical colonoscopy videos show that the reconstructed lumen geometry exhibits anatomically plausible diameter ranges, indicating that the recovered scale is stable and physically interpretable under known camera intrinsics. Overall, the proposed framework provides a sensor-free way to obtain scale-consistent depth and motion under a known imaging model, enabling more reliable geometry-aware reconstruction and downstream analysis in monocular colonoscopy.