3D Lung Reconstruction from X-Rays: Implicit Single-View Occupancy Vs. Explicit Voxel-Based Biplanar Methods

Deep learning has enabled 3D anatomical reconstruction from 2D radiographs through two main paradigms: image-to-segmentation and image-to-mesh methods. Yet, their relative performance remains poorly understood for the same anatomy. This study compares KVNet, a biplanar segmentation-based network, with the occupancybased model X2V, the state-of-the-art single-view implicit image-to-mesh model for lung reconstruction, on the same dataset of 646 cases. KV-Net achieves higher volumetric and surface accuracy, while X2V produces plausible global shapes with lower computational costs. Beyond whole-organ metrics, a regional analysis highlights distinct challenges across lung zones: occupancy-based models struggle to capture thin lower regions, whereas segmentation networks achieve a twofold reduction in average surface error across all regions. These results provide the first region-aware comparison between image-to-segmentation and image-to-mesh approaches for 3D lung reconstruction from X-rays.