Uncertainty-aware multi-stream deep learning for pediatric orthopedic bone age assessment from hand radiographs

Pediatric bone age assessment (BAA) is a fundamental task in pediatric orthopedics, endocrinology, and radiology, used to evaluate skeletal maturity in children and adolescents by comparing ossification patterns in left-hand radiographs against age-matched norms. In orthopedic practice, accurate BAA is indispensable for surgical planning in limb-length discrepancy correction, scoliosis management, epiphysiodesis timing, and predicting residual growth potential, all of which require reliable estimation of skeletal maturity to optimize intervention outcomes. Beyond orthopedic applications, automated BAA also supports the diagnosis of growth disorders, endocrine abnormalities, and constitutional delay of puberty. To more effectively exploit the heterogeneous information present in pediatric hand radiographs, we construct a three-stream fusion framework in which each branch is dedicated to a distinct radiographic representation derived from the RSNA Pediatric Bone Age dataset. The original radiograph is processed by an InceptionV3 backbone to encode global skeletal morphology and overall structural context, a soft-tissue–enhanced view is fed into a ResNet-50 backbone to emphasize cortical contours and soft-tissue boundaries, and a bone-masked view is analyzed by VGG-19 to focus on trabecular patterns and skeletal density. The high-level feature vectors from these three backbones are mapped into a common embedding space via lightweight 1$$\times $$1 projection layers, concatenated with a 32-dimensional gender embedding that accounts for sex-specific maturation trajectories, and subsequently passed to a linear regression head to produce a continuous bone-age estimate. Beyond the multi-stream framework, we introduce an uncertainty-aware Xception model that jointly predicts bone age and sample-wise aleatoric uncertainty, as well as an InceptionV3 model with bilinear pooling that captures second-order skeletal feature interactions critical for encoding subtle pediatric ossification cues such as carpal maturation and epiphyseal fusion. Our best single model achieves a mean absolute error (MAE) of 4.10 months on the RSNA dataset, outperforming all prior single-model methods without requiring additional annotations or ensemble strategies. This multi-stream design enables the model to aggregate complementary skeletal cues from structurally diverse inputs while maintaining a compact, end-to-end trainable architecture suitable for clinical deployment.