The study included two pediatric cohorts: cohort 1, comprising syndromic and endocrine disorders, and cohort 2, comprising lysosomal storage disorders. Radiographs were evaluated by human raters using the Greulich–Pyle method. In cohort 1, each radiograph was independently assessed by up to three raters, while in cohort 2, evaluations were performed by five raters. The panel of raters intentionally covered a broad spectrum of clinical expertise, ranging from medical students and residents in pediatrics to senior pediatric endocrinologists with more than 30 years of clinical experience. This design allowed for the assessment of interrater variability across different levels of training. In parallel, BA was also predicted by Deeplasia, a deep learning-based algorithm, which provided an independent reference estimate of skeletal maturity. For comparative analyses, patients were stratified by sex and further subdivided into four age groups based on chronological age at the time of imaging: 1) younger than 48 months (<4 years), 2) 48–96 months (4–8 years), 3) 96–144 months (8–12 years), and 4) older than 144 months (>12 years). This stratification was chosen to reflect clinically relevant developmental stages of childhood and adolescence, thereby facilitating both sex-specific and age-specific comparisons of bone maturation patterns across diagnostic groups. Because patient age at the time of image acquisition was not available for all images, complete age group assignment was not possible; therefore, all patients with unknown chronological age were assigned to an additional group 5.

Deeplasia (34) is a state-of-the-art prior-free deep convolutional neural network developed for the fully automated estimation of skeletal maturity from pediatric hand radiographs (34). In contrast to traditional atlas-based or explicit region-of-interest (ROI) approaches, the system utilizes an end-to-end learning strategy that enables it to capture both global and local image features in a data-driven manner, without relying on handcrafted priors. This design choice is critical for generalizing to atypical skeletal morphology, frequently encountered in patients with syndromic and dysplastic disorders. The core models were trained on the Radiological Society of North America (RSNA) Pediatric Bone Age dataset, which comprises over 14,000 radiographs, utilizing the Greulich–Pyle method as the consensus reference standard. The network architecture is based on EfficientNet variants. To account for known sex-specific maturation patterns, Deeplasia integrates sex information as a binary covariate, embedded and concatenated with the image features prior to the final regression. The system operates as a robust ensemble of three deep CNN variants derived from the described architecture. The final prediction is generated as the mean of these three independent model estimates. The process includes an initial automated hand-segmentation step to normalize image input and minimize the influence of artifacts, labels, or borders. As an open-source AI, the entire pipeline, including preprocessing masks and source code, has been made publicly available (34) to ensure transparency and reproducibility. Deeplasia has been previously validated on external clinical cohorts, demonstrating high test–retest precision suitable for longitudinal applications and high accuracy in patients with genetically confirmed skeletal dysplasias (32, 34).

Accuracy was assessed using the MAE, root mean squared error (RMSE), and the proportion of predictions within ±12 months (1-year accuracy). The relationship between Deeplasia and the reference standard was graphically assessed using scatter plots and Bland–Altman analysis (see Figures 1-3). Subgroup analyses were conducted based on the diagnosis, sex, and age group definitions, outlined in Section 2.1 (see Tables 1, 2). For cohort 1, while a complete review by three human raters was available for the subset of n = 849 images, the remaining 101 images were assessed by two raters, and all 950 images were included in the overall statistical analysis by using the mean BA of the available raters as the reference. To enable a fair comparison of the performance of Deeplasia with that of individual human raters, we employed a “leave-one-out” cross-validation procedure. Specifically, the mean BA assessed by the other experts served as the reference standard. We then computed the MAE, RMSE, and 1-year accuracy for both the excluded rater and Deeplasia, relative to this specific mean BA. This procedure was repeated for every individual rater in both cohorts (see Tables 3, 4). All images and their corresponding BA estimations were included in the statistical analysis. Outliers, as indicated in Figures 1–3, were defined as images for which the absolute difference between the Deeplasia prediction and the mean expert consensus BA exceeded 30 months. These BA estimations and predictions were subsequently examined individually (see Table 5).

The radiographic images from the first cohort of patients were analyzed by subdividing them into two primary groups based on their clinical presentation: disorders typically associated with accelerated skeletal maturation (see Figure 1) and disorders leading to delayed skeletal maturation (see Figure 2). In conditions with accelerated growth, Deeplasia demonstrated strong agreement with the expert consensus. Plot showed tight clustering along the line of identity, and Bland–Altman analysis revealed no systematic bias. In disorders with delayed growth, Deeplasia also showed strong agreement, although the mean difference line in the Bland–Altman analysis lay slightly above zero. In total, a minor positive bias of 1.52 months is recorded for all radiographs (see Table 1). Outliers, as shown in Figures 1 and 2, were defined as cases with a BA difference greater than 30 months between the mean expert evaluation and Deeplasia’s estimation. In total, four outliers were identified: one case of precocious puberty, one of SHOX deficiency, and two cases of Ullrich–Turner syndrome. The absolute BA differences in these cases ranged from 30.4 to 57.6 months. Three of these had been rated by only two raters, the remaining one by all three raters. Moreover, interrater disagreement was observed in these images, with standard deviations of expert evaluations ranging from 25.5 to 82.7 months (almost 7 years). This results in standard error of the means ranging from 18.0 to 58.5 months (see Table 5).

Table 1 summarizes Deeplasia’s performance across diagnostic, sex-, and age-specific subgroups in cohort 1. Across diagnostic categories, MAE values ranged from 5.44 to 9.10 months, RMSE values from 7.21 to 11.11 months, and 1-year accuracies from 70.0% to 92.47%. The lowest MAE values were observed in SHOX deficiency (5.44 months), Noonan syndrome (5.47 months), Silver–Russell syndrome (5.76 months), and congenital adrenal hyperplasia (5.76 months); the highest MAE was seen in pseudohypoparathyroidism (9.10 months). The highest STD of rater BA evaluations was also found for pseudohypoparathyroidism, with 11.88 months.

When stratified by sex, MAE was 6.04 months in female patients and 5.80 months in male patients. Deeplasia’s performance varied across age groups. The model performed best in early childhood (<4 years), with an MAE of 3.76 months, an RMSE of 5.1 months, and a 1-year accuracy of 96.1%. Performance decreased during mid-childhood and early puberty (8–12 years), reaching a maximum MAE of 6.78 months, a corresponding RMSE of 9.02 months, and a minimum 1-year accuracy of 85.8%, before improving again in adolescence (>12 years). Similar trends were observed for the 1-year accuracy and Pearson correlation coefficient. The bias showed an inverse pattern, with the lowest absolute bias of 0.82 months recorded in the 8–12-year age group.

Against individual raters (Table 3), Deeplasia achieved lower MAE and RMSE values and higher 1-year accuracies and Pearson correlation coefficients than any single human observer.

Figure 3 illustrates the agreement between Deeplasia and expert consensus in the lysosomal storage disorder cohort. The scatter plot show a strong linear relationship across the entire BA range, with data points closely aligned along the line of identity and no systematic deviation. The Bland–Altman plot and Table 2 reveal minimal bias of +1.04 months. Only two outliers were identified in cases of MPS I (BA difference −32.47 months) and MPS II (BA difference −35.80 months). For both cases, Deeplasia’s estimation was outside the range of the five expert evaluations (see Table 5).

Table 2 summarizes Deeplasia’s performance across diagnostic, sex-, and age-specific subgroups in cohort 2. MAE values across diagnostic categories ranged from 3.87 to 9.53 months, RMSE values from 4.70 to 13.90 months, and 1-year accuracies from 63.16% to 100%. The lowest MAE and the highest accuracy were observed in MPS III (MAE 3.87), while MPS II showed the largest deviations (MAE 9.53).

When stratified by sex, MAE was 5.81 months in female patients and 7.90 months in male patients.

Across age groups, the best performance was observed in the youngest children (0–4 years), with an MAE of 5.22 months, an RMSE of 6.68 months, and a 1-year accuracy of 91.3%. After this age range, both MAE and RMSE increased, reaching their peak in children aged 8–12 years, with an MAE of 9.74 months, an RMSE of 13.88 months, and a corresponding minimum 1-year accuracy of 69.2%, before improving again in older age groups.

Table 4 presents the comparison between Deeplasia and individual human raters in the lysosomal storage disorder cohort. Across all raters, Deeplasia achieved lower MAE and RMSE values and higher 1-year accuracies and Pearson correlation coefficients than any single observer.

The subgroup analysis of cohort 1 (Table 1) reveals that Deeplasia maintains high and consistent accuracy across most of the syndromic and endocrine disorders, with the MAE tightly clustered in a narrow range of 5.44 to 6.50 months. There was no systematic difference in its performance between conditions characterized by accelerated or delayed skeletal maturation. The only significant deviation was the elevated MAE in pseudohypoparathyroidism (9.10 months), which is most likely attributable to the marked and specific skeletal atypia associated with this rare condition (35). Crucially, even this peak MAE value still represents a very high level of accuracy, confirming that the average prediction error remains well below 12 months. The statistical findings are visually corroborated by the Bland–Altman analysis (Figure 1B). This finding suggests that determining bone age based on the Greulich and Pyle atlas is particularly challenging in such cases, as reflected by the fact that this disorder also exhibited the highest standard deviation among rater bone age evaluations within this cohort. Notably, the model achieved excellent results in the remaining dysplastic diseases, despite the known presence of pronounced hand dysplasias in many of these disorders (36, 37).

The most striking observation in the age stratification was a U-shaped accuracy curve. This pattern mirrors clinical experience of human BA evaluation, as these transitional growth phases are methodologically the most challenging to assess.

Furthermore, sex stratification yielded balanced results, confirming the model’s robust performance irrespective of gender. The slightly worse performance for female patients for this cohort is mostly correlated to the large sample size of only female Ullrich–Turner syndrome patients (n = 241), which was more difficult to assess (MAE = 6.5 months), while the mean MAE of all eight classes is slightly smaller, with 6.24 months (Table 1).

The four outliers in cohort 1 (see Table 5), where Deeplasia’s error relative to the mean of the raters exceeded 30 months, are most likely attributable to failures in the human rating process. These could include errors in data entry (e.g., typographical mistakes) or other sources of human error during rating, which led to unusually high standard deviations among rater evaluations and SEMs for these cases. Given the large number of images evaluated in this cohort (n = 950), such extreme deviations are expected and do not undermine the overall robustness of the study.

A fast and reliable bone age assessment is highly valuable for the conditions in cohort 1 for several reasons. It can serve as a diagnostic tool, for example, to detect early skeletal maturation and predict adult height in cases of precocious puberty (38), or to identify a typical bone age delay, as seen in Noonan syndrome (37). Another important application is the estimation of remaining growth potential, as well as therapy monitoring during growth hormone treatment in SHOX deficiency (39, 40), Noonan syndrome (41), Silver–Russell syndrome (42), and Turner syndrome (43). In congenital adrenal hyperplasia, treatment with glucocorticoids may accelerate skeletal maturation, which needs to be monitored to optimize growth outcomes (44). Across all these cases, an improved, standardized, and accelerated bone age assessment could significantly enhance patient management and treatment planning.

Performance across the diverse diagnostic subgroups in cohort 2 (LSDs) reflected the significant impact of phenotypic variability on model accuracy. The higher accuracy observed for MPS III (Sanfilippo syndrome) can be attributed to its relatively mild skeletal involvement and preserved bone proportions, which remain closer to the standard Greulich–Pyle reference morphology (19, 45). In sharp contrast, the lower accuracy for MPS II (Hunter syndrome) is consistent with its pronounced dysostosis multiplex and greater intradiagnostic heterogeneity, which substantially deviates from reference images (33). Similarly, the unclassified LSD group exhibited reduced precision, likely reflecting the broad radiographic and etiologic diversity within that category. Accuracy also showed a decline with age, a pattern associated with the lesser number of available images from older patients and the intrinsic difficulty of evaluating skeletal maturity in later stages of growth when deformities become more pronounced (33). The slight performance differences between sexes may also be explained by the genetic distribution of the included disorders: MPS II is inherited in an X-linked manner and therefore occurs almost exclusively in male patients (46, 47), while all other included lysosomal storage diseases, such as MPS I, III, IV, VI, and alpha mannosidosis (45, 48–52), are autosomal recessive. As a result, the male subgroup contained a higher proportion of severely affected MPS II patients with pronounced skeletal deformities, whereas the female subgroup included a less heterogeneous and overall milder spectrum of diseases.

The outliers in cohort 2 represented instances of disagreement between Deeplasia and all five raters (Table 5). However, since Deeplasia overall achieved higher accuracy than any individual rater when compared to the mean of the other raters’ evaluations, this disagreement cannot be considered systematic (Table 4). These findings confirm that Deeplasia maintains consistent accuracy and reproducibility even under the challenging skeletal conditions typical of lysosomal storage disorders.

A primary constraint is the geographic and ethnic homogeneity of the investigated cohorts, which predominantly reflect the demographic profile of Central Europe (Germany). While the AI model was trained on the globally used RSNA dataset, the generalizability of our specific validation findings to non-European or mixed-ethnicity populations remains limited, suggesting a need for future geographically diverse validation studies (23). Furthermore, the reliance on the Greulich–Pyle method (19) as the reference standard means the AI system inherits the inherent subjectivity and variability of this atlas-based approach, especially in transitional growth phases and in patients with pronounced dysostosis multiplex. Crucially, the absence of a true, independent ground truth means the consensus of human raters, inherently prone to variability, must serve as the reference for the AI system’s performance. Methodological limitations also include the uneven distribution and small sample size in certain rare diagnostic subgroups (e.g., precocious pseudopuberty and MPS VI). These small numbers increase the uncertainty of error metrics in these specific populations and may explain the observed performance drop compared to more prevalent disorders. Additionally, the multicenter nature of the data collection, spanning several decades, introduces unavoidable heterogeneity in image acquisition protocols. In particular, the inclusion of older radiographs in cohort 2 often resulted in image quality and resolution significantly inferior to modern clinical standards, which may attenuate the system’s and the rater’s performance compared to uniformly acquired datasets.

To reduce the impact of these conditions on the total statistics of a cohort, we tried to address this by reporting the mean statistical metrics—MAE, RMSE, 1-year accuracy, and bias—for all conditions in each cohort without disproportionately weighting any condition due to differences in sample size. Finally, as with all artificial intelligence models, there is a lack of explainability in the model’s assessments.