RSNA Bone Age Challenge Dataset

Updated 5 December 2025

RSNA Bone Age Challenge Dataset is a large, publicly available collection of pediatric hand X-rays with expert bone age and sex labels, serving as a benchmark for automated skeletal maturity estimation.
It supports diverse deep learning frameworks including supervised regression, attention-based localization, and ensemble methods to achieve high accuracy in bone age prediction.
The dataset has led to significant methodological advances with models reporting mean absolute errors below 4 months, despite challenges like class imbalance and label noise.

The RSNA Bone Age Challenge Dataset is a large, publicly released collection of pediatric hand radiographs with corresponding bone age and sex labels, designed as a benchmark for automated bone age assessment (BAA) in children. Established by the Radiological Society of North America (RSNA) as part of the 2017 RSNA Pediatric Bone Age Challenge, this dataset has become the de facto standard for evaluating machine learning approaches in medical image-based skeletal maturity estimation. The dataset's clinical-grade annotations, large scale, and open access have driven substantial methodological innovation in deep learning for medical imaging, with reported mean absolute errors (MAE) in months on held-out test sets now below 4 months in top-performing models (Kasani et al., 23 May 2024).

1. Dataset Composition and Labeling Protocol

The RSNA Bone Age Challenge Dataset comprises 14,236 digital postero-anterior (PA) X-ray images of pediatric left hands, spanning the chronological age range of 1 month to 19 years. Each image is annotated with:

Bone age: Defined as the chronological age in months, determined from patient date of birth and examination date and reviewed by a panel of board-certified pediatric radiologists.
Sex: Binary label (male/female), supplied for each subject.

The original data distribution reflects real-world clinical demographics, with few samples below 12 months or above 18 years and a predominance of cases in the 120–180 month (10–15 year) range. The RSNA challenge provided 12,611 images for training, 1,425 for development/validation, and an official held-out test set of 200 cases, with exact age and sex matching across splits (Kasani et al., 23 May 2024, Iglovikov et al., 2017).

2. Preprocessing and Annotation Strategies

Preprocessing pipelines in published studies address several unique challenges:

Image Quality and Heterogeneity: The raw RSNA images vary in resolution (up to 2080×1600 px), contrast, and background artifacts. Standard procedures include intensity normalization, histogram equalization, and foreground segmentation to isolate the hand (Iglovikov et al., 2017, Kasani et al., 23 May 2024).
Hand Segmentation: U-Net or Deeplab V3+ architectures are typically employed to delineate the hand, with initial training sets built from manually annotated masks (e.g., 100–751 hand-labeled images). Iterative “positive mining” and morphological post-processing further refine the segmentation (Iglovikov et al., 2017, Kasani et al., 23 May 2024).
Registration and Orientation Normalization: Key-point detection models (e.g., VGG-style CNN) align anatomical landmarks—the tip of the middle finger, capitate center, and thumb tip—enabling consistent cropping and correcting for handedness and rotation (Iglovikov et al., 2017, Kasani et al., 23 May 2024).
ROI Extraction: Inspired by clinical scoring (e.g., Tanner-Whitehouse method), regions-of-interest (ROIs) frequently include the whole hand, carpal bones, metacarpal region, and selected phalanges, with cropping coordinates derived post-registration (Iglovikov et al., 2017, Kasani et al., 23 May 2024).

A summary of dataset statistics and typical pipelines is provided below:

Statistic	Value	Notes
Images (total)	14,236	12,611 train, 1,425 val, 200 test
Age range	1–228 months (0–19 years)	Skewed toward 120–180 months
Sex ratio	~54% male/46% female in main splits	Matched in held-out test set

3. Learning Frameworks and Downstream Tasks

The dataset supports multiple learning paradigms:

Supervised Regression and Classification: Models predict continuous age or discrete age bins, often formulating the problem as regression or ordinal classification (Iglovikov et al., 2017, Wu et al., 2018, Kasani et al., 23 May 2024). Early approaches used VGG, ResNet, Inception, and custom architectures, often in ensembles.
Attention and Localization Mechanisms: Hierarchical residual attention modules (Wu et al., 2018), spatial part-relation frameworks (Ji et al., 2019), and class activation maps (Chen et al., 2020) allow models to focus on clinically relevant ROIs without external annotation.
Unsupervised and Semi-supervised Methods: CCAE-based architectures perform clustering of manually extracted ROIs with no labels, achieving ~76% accuracy in 48-month age-group classification (Zhu et al., 2022).
Divide-and-Conquer Strategies: Multi-branch architectures process several small hand patches (e.g., middle finger, carpal block, thumb base) independently, yielding enhanced MAE with lightweight models (Kasani et al., 23 May 2024).

Notable training details across works include aggressive online data augmentation, on-the-fly cropping/resizing, and explicit inclusion of gender as an additional feature (Wu et al., 2018, Kasani et al., 23 May 2024).

4. Performance Benchmarks and Comparative Evaluation

Recent work documents rapid improvement in reported accuracy. Key results (MAE in months, lower is better) include:

RSNA Challenge Winner (Iglovikov et al.): 4.97 (official test set, ensemble, three-region model) (Iglovikov et al., 2017)
PRSNet: Single model 4.49, ensemble <4.49 on the 200-image RSNA test set (Ji et al., 2019)
Attention-guided region learning: 4.3 (single model, no manual labels or bounding boxes, held-out test) (Chen et al., 2020)
Lightweight divide-and-conquer MobileNetV2 ensemble: 3.90 for 0–20 yr, 3.84 for 1–18 yr (RSNA official test set) (Kasani et al., 23 May 2024)

Supervised models continuously outperform unsupervised approaches (e.g., BA-CCAE, 76% accuracy for coarse age bins) (Zhu et al., 2022). Incorporation of domain knowledge—such as ossification patterns and age-specific ROI selection—yields further gains.

A representative comparison table:

Method	Test MAE (months)	Supervision
VGG-16 (baseline)	14.0	Supervised
U-Net+VGG (whole hand)	8.08	Supervised
Iglovikov et al. (ensemble)	4.97	Supervised (regions)
PRSNet (ensemble)	<4.49	Supervised (part-based)
Attention-guided distillation (agg)	4.3	Supervised (no ROI anno)
MobileNetV2 D&C (ensemble)	3.90	Supervised, lightweight
BA-CCAE (unsupervised, 48-mo groups)	~76% accuracy	Unsupervised (clustering)

5. Dataset Limitations and Technical Challenges

Several technical and clinical limitations impact modeling:

Class Imbalance: Severe scarcity of hand X-rays below 12 months or above 18 years leads to under-representation and increased prediction error outside the 1–18 year core bracket (Kasani et al., 23 May 2024, Iglovikov et al., 2017). Targeted augmentation partially mitigates but does not eradicate this bias.
Imprecision in Labeling: Slightly subjective clinical bone-age assignment, especially at developmental boundaries, introduces label noise, a challenge for fine-grained regression (Iglovikov et al., 2017).
Image Artifacts and Heterogeneity: Variations in image orientation, scanner type, background noise, and annotation tags necessitate robust preprocessing, including hand segmentation, background removal, and rotation normalization (Kasani et al., 23 May 2024).
ROI Annotation Burden: While masked and landmarked datasets boost accuracy, initial annotation effort is substantial and non-scalable for broader populations (Iglovikov et al., 2017, Zhu et al., 2022).
Reproducibility Gaps: Several studies do not report segmentation IoU/Dice, batch size, backbone architecture, or epoch count in sufficient detail for full replication (Wu et al., 2018).

6. Applications, Impact, and Future Directions

The RSNA Bone Age Challenge Dataset catalyzed a methodological shift away from classical computer vision toward deep learning and attention-based architectures in BAA. Its multi-institutional diversity and curation by expert panels ensure clinical utility and facilitate generalization studies.

A number of future research trajectories are suggested by ongoing model performance ceilings and dataset limitations:

Augmentation of low-frequency age groups using GANs or super-resolution techniques to enrich data below 12 months (Kasani et al., 23 May 2024).
Replacement of manual ROI annotation with automatic segmentation/localization pipelines to enable scale-out to other anatomical sites or populations (Zhu et al., 2022).
Longitudinal studies and temporal modeling of skeletal development, which are underexplored due to the cross-sectional nature of the dataset but highlighted as potential directions (Kasani et al., 23 May 2024).
Integration of label distribution learning and ordinal regression to better capture age continuity and compensate for ambiguous labeling (Chen et al., 2020).
Hybrid semi-supervised methods utilizing small labeled sets to bridge the remaining gap to fully unsupervised pipelines (Zhu et al., 2022).

The dataset continues to drive technical advances, support rigorous benchmarking, and foster innovation in pediatric and general medical imaging for developmental assessment.

7. Significance in the Broader Research Landscape

The RSNA Bone Age Challenge Dataset stands as the primary benchmark in open-access, expert-curated pediatric BAA, enabling standardized comparison across advances in computer vision, attention mechanisms, model ensembling, and unsupervised representation learning. Its clinical-grade annotation, scale, and diversity have made it foundational for reproducible research, facilitating methodological advances that extend beyond BAA to general medical image analysis, including segmentation, registration, and model interpretability (Iglovikov et al., 2017, Wu et al., 2018, Kasani et al., 23 May 2024).