Triple-level domain adaptation for bronchoscopic depth estimation with enhanced anatomical fidelity

Monocular depth estimation is a key technology in bronchoscopy navigation systems, providing essential information for 3D reconstruction and precise localization. However, the limited availability of annotated data hinders the application of supervised learning methods for monocular depth estimation. To address this limitation, unsupervised domain adaptation (UDA) has been introduced to leverage labeled source-domain data and improve target-domain performance. Existing single-level UDA methods can reduce domain discrepancies but often distort the spatial structure of bronchial lumens or fail to preserve fine-grained local details, diminishing both the anatomical fidelity and geometric consistency in the predicted depth maps. To overcome these challenges, we introduce triple-level domain adaptation for depth estimation (TriDADE), which unifies image-level adaptive Fourier domain adaptation, feature-level domain adaptation, and pixel-level discrimination with consistency constraints, to refine depth maps jointly. Experiments demonstrate that TriDADE achieves the lowest absolute relative error of 0.0971, surpassing even the intra-domain training performance, while improving depth prediction accuracy under the 1.25 threshold ([Formula: see text]) to 92%, significantly outperforming existing UDA methods. Qualitative results further elucidate the advantages of preserving bronchial lumen boundaries and fine anatomical details while producing smooth and geometrically consistent depth gradients. The proposed TriDADE framework effectively addresses the limitations of existing UDA methods, delivering substantial gains in both quantitative accuracy and anatomical fidelity. These improvements support the development of downstream clinical applications such as bronchoscopy navigation, 3D reconstruction, and precise diagnosis.