Bone age assessment using general use convolutional neural networks

Abstract

Deep Learning methods have been applied to different medical imaging analysis tasks like, e.g., lesion classification and tissue segmentation. Bone age assessment is traditionally performed on an x-ray of the non-dominant hand applying the Greulich and Pyle or the Tanner Whitehouse methods. In this work, we first have tested several state-of-the-art Convolution Neural Networks models for assessing bone age that previously has shown great results in general computer vision tasks. Based on these results, we have developed/optimized a new model, which is presented here. For this purpose, we used transfer learning methods and trained the selected networks from scratch achieving a 7.89-month error rate when assessing bone age in females.