Bone2Gene: Next-generation Phenotyping of Rare Bone Diseases

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.