Bone Age Assessment Using Deep Learning on Segmented Ossified Tissues of One Finger Radiograph

Purpose: Bone age assessment based on manual analysis of hand radiographs is highly subjective and time-consuming. In this study a Convolutional Neural Networks (CNN) was developed for automatic bone age estimation based on the index finger only.

Methods: The study included 1501 hand radiographs (896 females aged 2-8 and 605 males aged 3-9). The ground truth bone age was determined by the Greulich and Pyle standard. A pre-processing algorithm was first developed to segment the index finger in the radiograph, align it vertically and segment the ossified tissues in this finger. Three databases were then generated: The original radiographs of the index finger, binary masks of the ossified tissues and grayscale image of the segmented ossified tissues. A regression deep learning algorithm was applied to assess bone age, using each database separately. The training dataset was augmented by modifying the tilt angle and dimensions of the fingers.

Results: The database of the binary masks of the ossified tissues yielded the highest correlation with the ground truth bone age. For females, this database resulted in a standard deviation of 8.01 months from the ground truth, while 85% of the cases were correctly identified within a one-year error. For males, a standard deviation of 9.02 months from the ground truth was obtained while 83.7% of the cases were correctly identified within a one-year error. Using the database of the original finger radiographs resulted in a significantly higher standard deviation - 9.8 months for females and 10.97 for males.

Conclusion: Bone age can be estimated by analyzing radiographs of the index finger only. Higher accuracy was obtained following segmentation of the ossified tissues. Accurate assessment using a radiograph of one finger will reduce patient exposure to ionizing radiation and improve patient care and management.