Deep Learning-Based Pediatric Bone Age Estimation Using Enhanced Images and Pretrained Models

Abstract— Pediatric specialists usually apply radiography images to evaluate the maturity of bones that large discrepancy between the evaluated age and chronological age suggests a growth disorder. This procedure usually takes time and is subject to intra-observer and inter-observer variability. Therefore, automated bone age assessment methods, especially deep learning-based methods, are essential for accurate and efficient evaluation of bone maturity.
This study uses a convolutional neural network for pediatric bone age estimation, where pre-trained models are used for transfer learning. Prior to fine-tuning the pre-trained model, the input images are preprocessed to enhance the poor quality of images. Additionally, the gender data is incorporated into the model to improve its outcome.
The presented method was evaluated on the Radiological Society of North America (RSNA) pediatric bone age dataset. The results show that fine-tuned Xception pre-trained model satisfactorily outperforms other pre-trained models with a mean absolute error (MAE) of ۹.۳ months, which is comparable to cutting-edge techniques. These results show that preprocessing and transfer learning can effectively enhance the prediction performance of the proposed method. Furthermore, adding gender data into our model further improved its accuracy, highlighting the importance of considering this factor in pediatric bone age estimation. These findings have the potential to help pediatric experts evaluate bone maturity quickly and reliably, thereby enhancing patient care.

Keywords:

Artificial Intelligence , Bone Age , Deep Learning , Transfer Learning