Evaluating Children’s Growth Based on Carpal Bones in Left-Hand X-Ray Images Using Deep Learning

Technical Summary: Evaluating Children's Growth Based on Carpal Bones in Left-Hand X-Ray Images Using Deep Learning

Problem Statement
Bone Age Assessment (BAA) is a critical diagnostic tool in pediatric radiology and endocrinology for evaluating skeletal maturation and guiding treatment for growth disorders. Traditional clinical methods, such as the Greulich–Pyle (GP) atlas and Tanner–Whitehouse (TW) scoring system, rely on manual visual comparison of ossification centers. These approaches are inherently subjective, time-intensive, and prone to inter- and intra-observer variability. While recent deep learning advancements have automated BAA, most state-of-the-art models treat the entire hand radiograph as a global input. Consequently, the specific predictive contribution of distinct anatomical regions—particularly the carpal bones beyond early childhood—remains under-characterized. This study investigates whether carpal bones alone contain sufficient age-related information to support accurate BAA across childhood and adolescence, potentially offering a more focused and efficient alternative to whole-hand analysis.

Methodology
The research utilizes the RSNA Pediatric Bone Age dataset, comprising 12,811 left-hand X-rays from pediatric patients (ages 1–228 months). Due to the imbalance in the original dataset (specifically the scarcity of samples in early infancy and post-puberty) and the lack of public labels for the official test set, the authors re-distributed the labeled training data into an 80:10:10 split for training, validation, and testing.

• Region of Interest (ROI): Unlike previous studies that analyze the whole hand or multiple regions, this work isolates the carpal bones. The upper 50% of each X-ray is cropped to preserve the carpal region, while the lower 10% is padded with white pixels to prevent truncation during rotation augmentation.
• Preprocessing Pipelines: Two distinct preprocessing pipelines were developed and compared:
  1. Pipeline 1 (Non-thresholded): Applies Contrast Limited Adaptive Histogram Equalization (CLAHE) for local contrast enhancement, followed by Gaussian blurring and median filtering for noise suppression.
  2. Pipeline 2 (Thresholded): Includes all steps from Pipeline 1 but adds an adaptive thresholding step to convert images into binary representations.
  • Note: A grid search determined optimal parameters (Gaussian kernel=11, median kernel=3, block size=33, C=2) based on manual validation of carpal bone contour clarity.
• Model Architecture: A ResNet-50 backbone, pre-trained on ImageNet, was adapted for regression by replacing the classification head with a fully connected layer to predict age in months.
• Training Strategy: A progressive fine-tuning approach was employed. Initially, only the regression head was trained. Subsequently, deeper layers were gradually unfrozen. The non-thresholded pipeline required full network fine-tuning to achieve optimal performance, whereas the thresholded pipeline stabilized after unfreezing the final 30 layers.
• Data Augmentation: During training, images were normalized and subjected to small rotations (±5°), horizontal/vertical shifts (up to 10%/5%), and zoom variations (5%).

Key Results

• Overall Performance: The best-performing configuration (Pipeline 1, without adaptive thresholding) achieved a Mean Absolute Error (MAE) of 8.38 ± 0.39 months on the test set. This result outperformed the thresholded configuration, which yielded a significantly higher MAE of 14.85 ± 0.72 months.
• Comparison with Prior Work: While the 8.38 months MAE is higher than the top-performing whole-hand or multi-region models (e.g., Kasani and Sajedi 2023 reported 3.84 months using MobileNetV2 on full hands), this study demonstrates that a model relying solely on the carpal region can achieve competitive accuracy relative to full-hand methods, particularly when compared to earlier ResNet-50 benchmarks on the RSNA dataset (e.g., Iglovikov et al. 2018 reported ~8.42 months for full-hand).
• Preprocessing Impact: The removal of the adaptive thresholding step was the most critical factor in performance improvement. The authors attribute the degradation in the thresholded model to the suppression of fine-grained textural cues essential for age estimation.
• Demographic and Age-Specific Performance:
  • Sex: The model performed slightly better on males (MAE 7.91 months) than females (MAE 8.95 months).
  • Age Groups: Accuracy varied by developmental stage. The model showed lower errors during early, mid, and late puberty (e.g., 6.27 months for late puberty males) where carpal ossification is active. Performance was reduced in early infancy, a limitation the authors attribute to dataset imbalance (only 13 samples for ages 1–10 months) and limited anatomical variability in partially developed ossification centers.

Significance and Claims
The paper claims that carpal bones encode substantial age-related information sufficient for reliable bone age prediction when combined with appropriate deep learning models and careful preprocessing. The study challenges the assumption that whole-hand analysis is strictly necessary for high-accuracy automated BAA.

The authors position their findings as a step toward more efficient and interpretable clinical systems. By focusing on a smaller, well-defined anatomical region (the carpal bones), such approaches could potentially reduce annotation effort, shorten evaluation time, and provide consistent measurements without the complexity of analyzing the entire hand. However, the authors remain modest regarding the absolute accuracy, acknowledging that while carpal-only models are viable, they currently do not surpass the performance of the most advanced whole-hand or multi-region hierarchical models. The study emphasizes that the choice of preprocessing—specifically avoiding aggressive binarization—is a decisive factor in maximizing deep learning performance for this medical imaging task.