Geometry-optimized chamfer loss for femur 3D reconstruction with distance-enhanced CT segmentation

Abstract

Window-width thresholding serves as a common approach for isolating regions of interest from CT scans. In practice, however, it often produces spurious structures and cannot adequately suppress background noise. This paper proposes an adaptive loss formulation that dynamically adjusts target weights to improve segmentation fidelity. For epiphyseal regions—where bone boundaries lose sharpness—we combine a distance-map-derived label with a multi-scale feature aggregation module inside the U2Net backbone, sharpening edge predictions. Following segmentation, the stack of binary images is turned into a point cloud. We then refine the chamfer distance metric to exploit the femur’s characteristic morphology, which leads to higher reconstruction quality and an accurate three-dimensional model.