Segmentation of Pediatric Hand Radiograph Using UNet for Bone Aging

Abstract

The skeletal growth of children is often assessed by calculating bone age. Often developmental age differs from chronological age; hence bone aging is one of the essential steps in the clinical procedure of estimating the biological maturity of children. Assessment of bone age can be done in a traditional, manual way or using automated methods. Manual methods like Greulich-Pyle (GP) and Tanner-Whitehouse (TW) are time-consuming and involve intra- and inter-rater variability. Automated software are probably more reliable, and several are trained on a caucasian dataset given by the Radiological Society of North America (RSNA), consisting of 12,611 X-rays of boys and girls between the ages of 0 and 18 years. The GP method compares the patient’s full-hand radiograph to reference images in the GP atlas. However, in different ethnic groups, the maturation of different bone segments varies compared to Caucasian children having the same chronological age. To address these difficulties, our collaborators have recently developed a method called the ”Segmental GP rating” (Oza et al. 2023 submitted) which explicitly accounts for inter-segmental variability in the bones of the hand. We aim to develop an AI-based model that reproduces this novel method. Our bone age model was developed using the RSNA dataset. A UNet architecture was employed to segment four regions of interest (ROIs) from the X-rays. The image crops generated from segmented ROIs were subsequently used to train a DenseNet regression model for predicting bone age for the full-hand and three individual ROIs (short bones, carpals, and wrist). The final age of full-hand radiograph was taken as weighted sum of the predicted ages for the three ROIs. The weight values for each ROI were estimated using multivariate linear regression on a validation set. The obtained weight values for short bones, wrists, and carpals in boys were 0.57, 0.26, and 0.18, respectively. For girls, the weight values for the same ROIs were 0.69, 0.16, and 0.17, respectively. By combining the model’s age predictions for the three segments using these weights, we obtained a Mean Absolute Distance (MAD) of 6.7 months for boys and 7.4 months for girls on the validation set.