The automated Greulich and Pyle: a coming-of-age for segmental methods?

Abstract

The well-known Greulich and Pyle (GP) method of bone age assessment (BAA) relies on comparing a hand X-ray against templates of discrete maturity classes collected in an atlas. Automated methods have recently shown great success with BAA, especially using deep learning. In this perspective, we first review the success and limitations of various automated BAA methods. We then offer a novel hypothesis: When networks predict bone age that is not aligned with a GP reference class, it is not simply statistical error (although there is that as well); they are picking up nuances in the hand X-ray that lie "outside that class." In other words, trained networks predict distributions around classes. This raises a natural question: How can we further understand the reasons for a prediction to deviate from the nominal class age? We claim that segmental aging, that is, ratings based on characteristic bone groups can be used to qualify predictions. This so-called segmental GP method has excellent properties: It can not only help identify differential maturity in the hand but also provide a systematic way to extend the use of the current GP atlas to various other populations.

Keywords: Greulich and Pyle (GP); RSNA image share; bone aging; computer vision; personalizability.

Figure 1

Segmental rating on the RSNA image, 1547.png, a male hand rated to be 10.5 years (126 months) of age. Network prediction using the full hand image was 10.2 years (122.6 months). Each segment is boxed, and its predicted age is marked alongside. Notice the discrepancy in the segmental predictions—between 112 and 136 months—suggestive of a marked differential maturity in the hand. In Oza et al. (2023), manual segmental ratings were: short bones: 10.5 years (126 months); radius and ulna: 10 years (120 months); carpals: 10 years (120 months). Notice the carpals prediction matches the manual rating, the short bones are somewhat underestimated by the network, and the radius and ulna are relatively advanced.

Figure 2

Empirical cumulative distribution functions for bone age predicted from three models on all validation X-rays of age 120 months (boys). Model architectures include two ResNet's (50 and 152) and one DenseNet161, each trained on the full RSNA training data (sex: male); see Chapke (2023) for the general methodology of model construction. Note that each model is evaluated on validation set of X-rays labeled with age 120 months (n = 42). Among the three models, predictions from each network differ slightly in estimates of age. The ensemble ecdf is overlaid in bold; ensemble predictions are observed to align with the average of the three models.