Joint Region of Interest Detection and Bone Age Estimation from Radiograph of the Hand

Abstract

Skeletal bone age assessment is a conventional clinical method used to determine adolescent maturity in orthodontics, kinematics, pediatrics, forensic science, and other professions. Bone age is typically determined using the Greulich and Pyle (GP) or Tanner Whitehouse (TW2 or TW3) procedures from radio-graphs of the hand. Because GP and TW methods are subjective and rely on the practitioner's skill, they demand long analysis time, as well as experienced staff. This indicates the necessity for a fully automated approach to determine bone age. In this study, we proposed a robust model to jointly detect wrist area and estimate the bone age from radiograph data. To this end, we initially examined several state-of-the-art Convolution Neural Networks (CNN) models to specify bone age. Then, we proposed the hybrid model, involving two stages. In the first stage, the object detection module selects the key regions of input image to aid the bone age estimation. In the second stage, an attention mechanism based on Residual Neural Networks (ResNet) was included to increase the accuracy of bone age estimation within the classification phase. The mean absolute error for the hybrid model was 0.75 years compared to 1.03, 1.23, 1.08 years obtained from Resnet 18, Resnet 50, and Mobilenet, respectively. The proposed model for joint detection of wrist area and estimation of bone age exhibited superior performance over the conventional networks. This model could be used a decision support tool to reduce the workload/processing time in clinical practice. © 2022 IEEE.