A Multi-Segment Fusion Architecture for Bone Age Estimation: Comparative Backbone Analysis and External Validation in a Single-Center Mexican Clinical Cohort

One of the most challenging tasks in pediatric medicine is bone age estimation from hand radiographs. Traditional bone age estimation approaches show limited performance across different demographic groups, making diagnostics prone to misclassification of growth abnormalities. On the other hand, convolutional neural networks have demonstrated remarkable performance for many computer vision tasks, including bone age estimation. These models automatically extract and learn meaningful patterns capturing structural variations in bones. However, these models still present limited generalization capabilities, especially across different demographic populations, and are rarely evaluated beyond the public datasets on which they are trained. This paper presents a fusion architecture, F-DenseNet121, together with a systematic evaluation of its generalization across demographic populations. The employed methodology incorporates an anatomical segmentation strategy inspired by the Tanner–Whitehouse 3 (TW3) framework, combined with feature learning from the RSNA Pediatric Bone Age Challenge dataset. Instead of training a standalone convolutional model, the proposed methodology considers a convolutional network for each anatomical segment. This mechanism allows the model to learn localized skeletal patterns, reducing the influence of irrelevant structures. During training and validation with the RSNA dataset, the proposed F-DenseNet121 obtained a Mean Absolute Error (MAE) of 5.77 months during internal validation, a figure comparable to several reported convolutional models under their respective internal validation protocols. F-DenseNet121 was also evaluated using an independent 10.8% RSNA test validation subset, obtaining an MAE of 13.70 months, a result that remains substantially higher than published state-of-the-art benchmarks (4.2–6.2 months) and reveals a substantial generalization gap between internal validation and independent testing. To further examine this gap, external validation was performed using radiographs from Mexican patients. In this test, all convolutional models showed a significant performance difference between the public RSNA dataset and the clinical data from Mexican patients, with F-DenseNet121 and F-InceptionV3 achieving statistically indistinguishable external performance among the evaluated backbones. Rather than positioning these results as evidence of state-of-the-art accuracy, this study highlights that internal validation performance can substantially overestimate real-world reliability, and underscores the value of rigorous, multi-architecture comparison and external clinical validation for assessing the true applicability of automated bone age estimation systems.