An AI-powered solution for automated bone age assessment from hand X-ray images
Bone age assessment is a crucial diagnostic tool in pediatric medicine for evaluating skeletal maturity and diagnosing growth disorders. This project implements state-of-the-art deep learning architectures (DenseNet121 and ResNet34+CBAM) that automatically predict bone age from hand X-ray images, potentially reducing assessment time from hours to seconds while maintaining clinical accuracy.
- Growth Disorders: Early detection of growth hormone deficiencies
- Endocrine Evaluation: Assessment of pubertal development
- Treatment Planning: Monitoring response to growth treatments
- Clinical Efficiency: Standardized, reproducible assessments
🎯 Automated Bone Age Assessment
- Direct age prediction from cropped and enhanced hand X-rays
- Support for pediatric patients (0–19 years)
- Clinical-grade accuracy comparable to expert radiologists
🔧 Advanced Deep Learning Pipeline
- DenseNet121 and ResNet34 with Convolutional Block Attention Module (CBAM)
- Auxiliary gender input integrated with visual features
- Cosine learning rate scheduling with warm-up
- AdamW optimizer with weight decay regularization
- Huber loss for robust regression
📊 Comprehensive Analysis
- Detailed performance metrics on training, validation, and held-out test sets
- Visualization of prediction confidence
- Error analysis and statistical evaluation
- Direct comparison of DenseNet vs. Attention-augmented ResNet
🚀 Production Ready
- Jupyter notebook implementation for research and development
- Modular code structure for easy integration
- Comprehensive documentation and examples
The model is trained on the RSNA Pediatric Bone Age Challenge dataset:
- Training: 12,611 images
- Validation: 1,425 images
- Test: 200 images (held-out for final evaluation)
- Image Format: Grayscale, cropped to hand region, resized to 256×256
- Labels: Bone age in months + binary gender (0=female, 1=male)
- Preprocessing: MediaPipe hand detection, CLAHE contrast enhancement, grayscale conversion, normalization to [0,1]
Two main models were implemented:
-
DenseNet121
- Dense connectivity with 4 dense blocks (6, 12, 24, 16 layers)
- L2 regularization and dropout
- Gender processed via small MLP and concatenated to image features
- Final regression head with linear output
-
ResNet34 + CBAM
- Residual network with [3,4,6,3] block groups
- Convolutional Block Attention Module for channel + spatial attention
- Gender input fused with pooled image features
- Deep regression head with dropout regularization
- Input Shape: (256, 256, 1) grayscale
- Batch Size: 64
- Optimizer: AdamW with weight decay
- Loss Function: Huber loss (δ=15)
- Learning Rate: Cosine decay with warm-up (base 5e-3)
- Early Stopping: Based on validation MAE
- Epochs: Up to 60 (early stopped)
- MAE: 8.18 months
- RMSE: 11.49 months
- R²: 0.93
- Params: 7.1M
- MAE: 9.77 months
- RMSE: 13.01 months
- R²: 0.90
- Params: 22.6M
| Model | MAE (months) | RMSE | R² | Params |
|---|---|---|---|---|
| DenseNet121 | 8.18 | 11.49 | 0.93 | 7.1M |
| ResNet34+CBAM | 9.77 | 13.01 | 0.90 | 22.6M |
➡️ DenseNet121 consistently outperformed ResNet34+CBAM in accuracy and efficiency.
- Preprocessing: MediaPipe hand cropping, CLAHE, grayscale conversion, aspect-ratio preserving resize, normalization
- Data Pipeline: tf.data with caching, shuffling, batching (64), prefetching
- Regularization: L2 penalties + dropout
- Training: AdamW optimizer, cosine LR schedule with warm-up, Huber loss
- Evaluation Metrics: MAE, RMSE, R²
✅ DenseNet121 is the best tradeoff: accurate, efficient, and stable
✅ Attention helps but adds complexity without consistent gains
✅ Robust preprocessing pipeline improves results
- Integrate attention directly into DenseNet
- Use multimodal clinical metadata (height, weight, ethnicity)
- Explore uncertainty estimation for interpretability
- Transfer learning from related medical imaging tasks
This project is licensed under the MIT License - see the LICENSE file for details.