ORCA: An ensemble deep learning framework for automatic detection and deformity assessment for lower-limb radiographs of skeletal dysplasia

Skeletal dysplasia (SD) is a group of congenital disorders mostly caused by germline mutations and encompassing over 700 conditions, affecting millions of people worldwide. Patients often presented with skeletal deformities in the lower extremities, and orthopedic corrections are often needed. Pre-operative assessment of the deformities is critical in clinical decision-making process. Currently, such assessments are often manually performed by radiologists or physicians, with significant inter-rater variabilities and low efficiencies. Deep-learning based approaches developed recently using individual models to predict key-points and outlines of the human skeleton are gaining traction, but the limitations inherent to individual techniques and the low signal-to-noise ratio nature of SD radiographs posed major challenges in outcome stability and consistency. Here, we presented ORCA, an ensemble framework whereby a scalable number of models were incorporated to present averaged or majority-voted results. We deployed ORCA to detecting key points of the lower limbs (ankle, knee, and hip) and the outlines of the femur and tibia in a set of SD patients and controls in-housely retrieved from our hospital. The results show that ORCA outperforms individual models, in terms of multiple criteria of accuracies, stabilities and consistencies. Our work highlights the potential of ensemble deep learning in automatic orthopedic radiography for SD.