Title: Reimagining Bone Analysis: Enhanced Pose Estimation for X-Ray and Ultrasound Imaging

Accurately measuring the angles between skeletal structures is a standard procedure in medical image analysis, serving as a critical quantitative metric for both diagnostic assessment and treatment planning. By automating these measurements, healthcare providers can lower costs, save time, and enhance the reproducibility of their results. This study introduces a novel approach to automatic bone pose estimation that utilizes a learning-based framework to propose point candidates, followed by a line model to derive axis parameters.

To address the vulnerability of conventional line models, such as least squares, to outliers, the proposed method integrates false-positive reduction strategies alongside robust fitting techniques, including RANSAC and Hough transforms. These enhancements significantly boost the reliability of the angle estimations. The methodology was evaluated across three clinically significant pediatric tasks: assessing fracture fragments in both radiographs and ultrasound images, and evaluating developmental dysplasia of the hip via ultrasound using the Graf method.

The results demonstrate mean errors of $4.1^\circ$ for radiographic fracture assessment, $5.4^\circ$ for ultrasound fracture assessment, and $5.51^\circ$ for hip dysplasia evaluation. These figures not only fall within the range of expected clinical observer variability but also represent a substantial improvement over traditional landmark-based methods. For further research and application, the code and annotations related to fracture angle assessment in radiographs have been made publicly available on GitHub.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC