Bone2Gene: Next-generation Phenotyping of Rare Bone Diseases
Year: 2026
Type: preprint
Abstract: Background Diagnosing the over 700 known rare bone diseases (RBDs) is inherently challenging and often requires extensive time and multiple clinical visits. Effective treatment, particularly for RBDs with approved therapies, depends on early and precise identification of the specific RBD type. Image recognition artificial intelligence (AI) has the potential to significantly enhance diagnostic processes and improve patient outcomes. Many of these disorders cause characteristic skeletal changes, especially in the hands, and are associated with growth abnormalities. Consequently, affected children routinely undergo hand radiographs for bone age assessment, making these images a widely available yet underutilized diagnostic resource. Materials and Methods We retrospectively compiled 5,623 multi-institutional hand radiographs from 2,471 patients with 45 different RBDs and 1,382 unaffected controls. We trained two deep learning models: a binary classifier to differentiate between RBD and non-RBD hand radiographs, and a multi-class classifier covering ten RBDs (or RBD groups), using 5-fold cross-validation. Preprocessing included masking, normalization, and data augmentation. Additionally, we applied occlusion sensitivity mapping to visualize class-specific features and evaluated the learned representations through cosine-based retrieval and UMAP projections of the feature space. Results The affected versus unaffected classifier achieved a balanced accuracy of 85.5% on the test dataset. The ten-class classifier reached a balanced (top-1) accuracy of 76.6%, with top-3 accuracy exceeding 90%. Disorders with highly distinctive phenotypes, such as achondroplasia, achieved accuracies above 95%, whereas phenotypically overlapping disorders, such as ACAN- and SHOX-related short stature, were more frequently confused. Feature space analysis showed that validation samples clustered closely with their respective training distributions, supporting the consistency and generalizability of the learned embeddings. Conclusion This manuscript presents a proof of principle for the development of Bone2Gene , a next-generation phenotyping (NGP) tool for the detection and differential diagnosis of RBDs, currently based on hand radiographs. Ongoing efforts focus on expanding the dataset to include additional RBDs or RBD groups in the current multi-class classifier for differential diagnosis and to further evaluate its generalizability. The Bone2Gene study is open to collaboration.
Source: medRxiv
Authors: Eike Bolmer, Philipp Schmidt, Isabel Fischer, Sebastian Rassmann, Adele Ruder, Alexander Hustinx, Aron Kirchhoff, Christoph Beger, Kyra Skaf, Minu Fardipour, Tzung-Chien Hsieh, Alexandra Keller, Alberto De Rosa, Silvia Kalantari, Fabio Sirchia, Primož Kotnik, Mark Born, Benjamin D. Solomon, Rebekah L. Waikel, Tinatin Tkemaladze, Luka Abashishvili, Elene Melikidze, Anastasia Sukhiashvili, Megi Lartsuliani, Julian Nevado, Jair Tenorio, Julian Jürgens, Mona Lindschau, Christina Lampe, Shahida Moosa, Jean Tori Pantel, Larissa Mattern, Miriam Elbracht, Ho-Ming Luk, André Travessa, Julieta De Victor, Maryam Alhashim, Amal Alhashem, Noura AlKaabi, Sinem Kocagil, Emre Akbaş, Uwe Kornak, Tilman Rohrer, Roland Pfäffle, Ondřej Souček, Zdeněk Šumník, Andreas Zankl, Himanshu Goel, Markus Nöthen, Ulrike Attenberger, Amaka C. Offiah, Lothar Seefried, Hermann Josef Girschick, William Wilcox, Denny Schanze, Martin Zenker, Pablo Lapunzina, Peter Krawitz, Klaus Mohnike, Behnam Javanmardi
Institutions: University of Bonn, University Hospital Magdeburg, Otto-von-Guericke-Universität Magdeburg, German Center for Neurodegenerative Diseases, Leipzig University, University of Pavia, Policlinico San Matteo Fondazione, Ljubljana University Medical Centre, University Hospital Bonn, National Human Genome Research Institute, Tbilisi State Medical University, Hospital Universitario La Paz, Centre for Biomedical Network Research on Rare Diseases, University Medical Center Hamburg-Eppendorf, Universitätsklinikum Gießen und Marburg, Stellenbosch University, RWTH Aachen University, Hong Kong Children's Hospital, Centro Hospitalar Lisboa Norte, King Fahad Specialist Hospital, Eskişehir Osmangazi University, Universitätsmedizin Göttingen, Universitätsklinikum des Saarlandes, University Hospital Leipzig, Charles University, Fakultní nemocnice Motol a Homolka, The University of Sydney, Garvan Institute of Medical Research, Children's Hospital at Westmead, Hunter Genetics, University of Newcastle Australia, Medical University of Vienna, University of Sheffield, University of Würzburg, Klinikum im Friedrichshain, Emory University
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7141171023