Advancing Bone Age Estimation with Xception: A Deep Learning Model for Pediatric Radiograph Analysis

Bone age is one of the critical determinants of skeletal maturity in children, aiding in the diagnosis of disorders due to growth and dental development. Currently, assessment using the Greulich-Pyle Atlas is labor-intensive and inconsistent, since interpretations are subjective by manual evaluation. Based on this perspective, a deep learning-based methodology is presented which proposes the automation of the determination of bone age from radiographs of the hand and wrist through an Xception model. The authors of the study used 12,611 images from the RSNA Bone Age dataset resized to 256×256 pixels, as well as rotation, zoom, brightness adjustment, and horizontal flipping, in order to increase the variability of images and prevent overfitting. In this experiment, the authors used the Xception model, which extracts features very efficiently through depthwise separable convolutions, thus not burdening computational resources. After this step in the experiment, hyperparameter optimization was possible based on learning rate, units in the fully connected layer, and dropout rate by using the Hyperband algorithm. The model was then trained with the Adam optimizer, where it reached an MAE of 5.56 months on the test set, and its robustness toward external data came out as an MAE of 6.00 months. This highly efficient and accurate model can be applied clinically in order to expedite processes that take up much time and can be laden with variability from human observations.