A deep automated skeletal bone age assessment model via region-based convolutional neural network

Skeletal bone age assessment is widely applied in growth prediction and auxiliary diagnosis of medical problems. X-ray images of hands are observed in the evaluation of bone age, where the ossification centers of epiphysis and carpal bones are the key regions. Traditional skeletal bone age assessment methods extract these areas to predict the bone age but few of them can achieve satisfactory efficiency or accuracy. While automatic bone age assessment methods with deep learning techniques have achieved the leading performance, most of them can only accept fixed-size small images and ignore these key regions. In this paper, we take full consideration of the significant regions and propose a novel deep automated skeletal bone age assessment model via region-based convolutional neural network (R-CNN). We transfer Faster Region-based Convolutional Neural Network (Faster R-CNN) model from object detection to bone age regression in order to detect the ossification centers of epiphysis and carpal bones and evaluate bone age. The proposed model has overcome the limitation of other CNN based models, taking large-scale original X-ray images as inputs. It can automatically extract the features, detect the key regions and further predict the bone age. To validate the effectiveness of the proposed model, we realized different prior methods and conducted a series of experiments on two data sets using 10-fold cross-validation to compute the Mean absolute errors (MAEs). The results show that the MAEs of the proposed model are 0.51 and 0.48 years old respectively, better than other bone age assessment methods including state of the art.