enLLASD: An Ensemble Deep Learning Framework to Automate Derivation of Lower Limb Alignments for Skeletal Dysplasia

Skeletal dysplasia (SD) is a group of rare, congenital skeletal disorders that affect millions worldwide. Patients often present with moderate to severe limb deformities, which need to be constantly monitored radiographically, in a process called alignment assessments. Currently, alignments are usually determined manually by physicians, with significant inter-rater variability and low efficiencies. Recent efforts to automate this often used individual deep learning models to detect bone contours or landmarks (e.g., joints). Due to data scarcity, case heterogeneity, and low bone mass, existing methods risk model overfitting. We propose enLLASD, an ensemble learning framework that integrates key-point detection and bone segmentation results from multiple member models, to enhance robustness. Predictions are aggregated using averaging, majority voting, and logistic stacking, while false positives in segmentation are suppressed through consistency with key-point predictions. A spline-based resampling method is used to fit the medial axes of the femur and tibia, enabling the computation of key alignment angles. We validated our framework on a dataset of 1416 full-length radiographs from both SD and non-SD individuals. Experimental results demonstrate that logistic stacking [intersection over union (IOU) for front femur: 0.9564, front tibia: 0.9468] outperforms individual models in segmentation accuracy and alignment angle estimation, particularly in cases with severe deformities. Our work highlights the potential of ensemble deep learning in automatic orthopedic radiography for SD.