Automated segmentation and length measurement of metacarpal and phalangeal bones for hand radiograph evaluation.
Authors
Affiliations (6)
Affiliations (6)
- Institute for Genomic Statistics and Bioinformatics, Universitätsklinikum Bonn, Baunscheidtstr. 17, 53113, Bonn, Germany.
- Medical Faculty, Otto-Von-Guericke-University Magdeburg, Leipziger Straße 44, 39120, Magdeburg, Germany.
- Human-Centered Artificial Intelligence, University of Augsburg, Universitaetsstrasse 6a, 86159, Augsburg, Bavaria, Germany.
- b-it and Computer Science Department, University of Bonn, Bonn, 53115, Germany.
- Lamarr Institute for Machine Learning and Artificial Intelligence, Bonn, Germany.
- Institute for Genomic Statistics and Bioinformatics, Universitätsklinikum Bonn, Baunscheidtstr. 17, 53113, Bonn, Germany. [email protected].
Abstract
Evaluating hand and wrist radiographs is essential in pediatric endocrinology and clinical genetics, particularly for the assessment of suspected skeletal anomalies. In this study, we present Auto-Bone-Caliper, an automated system for the segmentation and length measurement of metacarpal and phalangeal (M&P) bones, trained and evaluated on public datasets comprising both normal and dysmorphic cases. We first introduce InstanceSAM, a two-stage framework that detects and segments all 19 M&P bones in pediatric hand radiographs, achieving Dice scores of 98.7% for normal bones and 95.0% for dysmorphic bones. We further develop and evaluate three methods for bone-length estimation, identifying a k-means-based approach as the most accurate, with relative errors of 2.2% for normal bones and 4.5% for dysmorphic bones. Our automated pipeline, Auto-Bone-Caliper, integrates InstanceSAM with the k-means-based length-estimation method. To enable scale-independent downstream analyses, we derive relative bone-length measures from the automated measurements. Using these relative measures, we statistically compare measurements obtained using Auto-Bone-Caliper on an independent dataset with a healthy reference catalog of normal bone morphologies, observing a high level of agreement (Wasserstein-1 distance = 0.012). Finally, we demonstrate a potential clinical use case of Auto-Bone-Caliper by obtaining relative metacarpophalangeal pattern profiles for three genetic conditions, namely Turner syndrome, achondroplasia, and pseudohypoparathyroidism. Our results highlight the potential of the Auto-Bone-Caliper to streamline and standardize M&P length measurement, providing an objective and reproducible tool suitable for clinical application.