Summary

This study developed a robust deep learning model for determination of skeletal age based on routinely available hand radiographs that can be integrated into clinical workflows for sports medicine physicians evaluating skeletal maturity.

Abstract

Background

Surgeons treating skeletally immature patients rely on assessment of physeal status to determine appropriate and safe surgical strategies (e.g. physeal-sparing anterior cruciate ligament reconstruction). The gold standard method of skeletal maturity estimation is the Greulich and Pyle atlas based on hand radiographs, which is notably time-consuming. Deep learning (DL) has previously successfully identified orthopedic implants with incredible accuracy. This study sought to develop a highly reliable DL model for determination of accurate skeletal age based on hand radiographs for routine use by orthopedic practitioners.

Methods

This study utilized three publicly available hand radiograph datasets for model development and validation: 1) the Radiological Society of North America (RNSA) dataset containing 12,611 x-rays, 2) the Radiological Hand Pose Estimation (RHPE) dataset containing 6,288 films, and 3) the digital hand atlas (DHA) comprising 1,400 radiographs. All three datasets report corresponding sex and skeletal age, and the RHPE/DHA datasets additionally contain chronological age. Image pre-processing was composed of three stages – a pre-trained U-net for hand mask prediction, a histogram equalization technique for image improvement through contrast, and alignment standardization. The ConvNeXt architecture was chosen as the background model for bone age estimation. First, the model was trained on the RSNA dataset, with corresponding sex/skeletal age as inputs. To improve model accuracy, we chose to incorporate chronological age given its utility in the clinical setting. Therefore, the model was then subsequently trained on the RHPE dataset, with final model validation performed on the DHA.

Results

The first model was trained for 100 epochs on the RSNA dataset with the Mean Squared Error (MSE) loss function, using early stopping in order not to overfit on the training set. This model achieved a mean absolute difference (MAD) of 6.4 months on the RSNA validation set, which outperformed the 6.9 months achieved by the “winning” model of the RSNA competition when externally validated. After incorporation of chronological age and further model enhancement on RHPE dataset, this error improved to an MAD of 6.2 months on the RHPE validation set, similarly, surpassing the previously best reported value of 6.3 months on this series of images. For final model validation and assessment of its generalizability, the model was applied to the DHA dataset. This achieved an MAD of 5.6 months, again surpassing the best RSNA model when applied to the DHA, which represents a diverse population of individuals.

Conclusions

Leveraging newer DL technologies trained on nearly 20,000 hand radiographs across three distinct, diverse datasets, this study developed a robust model for predicting bone age. Utilizing features extracted from the model trained on the RSNA dataset, combined with incorporation of chronological age provided in the RHPE dataset, our model outperforms prior state of the art models when applied to external validation datasets. These results indicate that our model provides a highly accurate platform for integration into clinical workflows that can affect real-time decision-making for sports medicine physicians, radiologists, and other musculoskeletal practitioners evaluating skeletal age.