Automated Bone Age Estimation Using Deep Learning

A deep learning-based system for automated pediatric bone age estimation using hand X-ray images with Xception architecture and transfer learning.

📋 Table of Contents

• Overview
• Key Features
• Performance Metrics
• Dataset
• Model Architecture
• Installation
• Usage
• Project Structure
• Results
• Team
• Acknowledgments

🎯 Overview

Bone age assessment is a critical diagnostic tool for evaluating skeletal maturity and diagnosing growth disorders in children. Traditional manual methods are time-consuming and subject to inter-observer variability (±6-12 months). This project implements an automated bone age prediction system using deep learning to achieve clinically acceptable accuracy.

Key Achievements:

• 📊 R² Score: 0.9169 (target: ≥0.90)
• 📉 MAE: 9.04 months (target: ≤12 months)
• ⚖️ Low Gender Bias: <2 months MAE difference between genders
• 🎯 91.54% Classification Accuracy for developmental stages

✨ Key Features

• Transfer Learning: Leverages Xception architecture pre-trained on ImageNet
• Gender-Aware Prediction: Incorporates patient sex as input for improved accuracy
• Model Explainability: Grad-CAM visualization shows anatomically-correct attention on carpal bones and growth plates
• Comprehensive Analysis:
  • Regression performance metrics
  • Gender bias analysis
  • Developmental stage classification (Child/Adolescent/Adult)
  • Error pattern analysis
• Mixed Precision Training: Optimized for RTX 3060 12GB GPU
• Robust Preprocessing: Architecture-specific preprocessing with data augmentation

📊 Performance Metrics

Regression Performance

Metric	Value
R² Score	0.9169
MAE (Mean Absolute Error)	9.04 months
RMSE	11.75 months
Within ±12 months	91.5%

Classification Performance

Metric	Value
Overall Accuracy	91.54%
Quadratic Weighted Kappa (QWK)	0.8248
Child (0-10y) Recall	95%
Adolescent (10-18y) Recall	91%

Gender Bias Analysis

• MAE Difference: <2 months between male/female samples
• R² Difference: <0.01
• Fair predictions across both genders ✓

📁 Dataset

RSNA Pediatric Bone Age Challenge Dataset

• Total Images: 12,611 hand X-ray images
• Age Range: 1-228 months (0-19 years)
• Modality: Left hand radiographs
• Metadata: Ground truth bone age + patient sex
• Split: 70% training / 15% validation / 15% test (stratified by age)

🏗️ Model Architecture

Xception Base Model

• Architecture: 36 convolutional layers with depthwise separable convolutions
• Pre-training: ImageNet weights
• Innovation: Efficient feature extraction through extreme Inception modules
• Training Strategy: All layers trainable from start (not frozen)

Regression Head

Xception Base (256×256×3 input)
    ↓
GlobalMaxPooling2D
    ↓
Flatten
    ↓
Dense(10, ReLU)
    ↓
Dense(1, linear) → Bone Age (months)

Training Configuration

• Input Size: 256×256 pixels
• Preprocessing: xception.preprocess_input() ([-1, 1] scaling)
• Batch Size: 4 (memory constraint)
• Optimizer: Adam (lr=0.001)
• Loss Function: Mean Squared Error (MSE)
• Early Stopping: Patience=7 epochs
• Epochs: 50 max (converged at epoch 18)
• Mixed Precision: FP16 for memory efficiency

🛠️ Installation

Prerequisites

Python 3.8+
CUDA 11.2+ (for GPU support)

Dependencies

pip install tensorflow>=2.8.0
pip install numpy pandas matplotlib seaborn
pip install scikit-learn opencv-python
pip install jupyter notebook

Clone Repository

git clone https://github.com/AmitAK1/bone-age-estimation.git
cd bone-age-estimation

🚀 Usage

Training the Model

# Open the main notebook
jupyter notebook Bone_Age_Prediction_Xception_FINAL.ipynb

# Or run directly
python train_model.py  # If you create a standalone script

Making Predictions

from tensorflow.keras.models import load_model
from tensorflow.keras.applications.xception import preprocess_input
import cv2

# Load model
model = load_model('best_model.h5')

# Load and preprocess image
img = cv2.imread('xray_image.png')
img = cv2.resize(img, (256, 256))
img = preprocess_input(img)
img = np.expand_dims(img, axis=0)

# Predict (with gender: 0=female, 1=male)
gender = 1  # male
prediction = model.predict([img, np.array([[gender]])])
print(f"Predicted Bone Age: {prediction[0][0]:.2f} months")

📂 Project Structure

bone-age-estimation/
│
├── Bone_Age_Prediction_Xception_FINAL.ipynb  # Main implementation notebook
├── README.md                                   # Project documentation
├── requirements.txt                            # Python dependencies
├── .gitignore                                  # Git ignore rules
│
├── Bone_age_prediction/                        # Output visualizations
│   ├── training_history.png                    # Training curves
│   ├── prediction_scatter.png                  # Predicted vs actual
│   ├── confusion_matrix.png                    # Classification results
│   └── gradcam_*.png                           # Grad-CAM visualizations
│
├── docs/                                       # Documentation files
│   ├── Bone_Age_Report_Fixed.tex              # LaTeX report source
│   ├── Bone_Age_Presentation.tex              # LaTeX presentation
│   ├── PRML_COURSE_PROJECT_Final.pdf          # Final report PDF
│   └── PRML_PROJECT_PRESENTATION.pdf          # Presentation slides
│
└── models/                                     # Saved model weights
    └── best_xception_model.h5                 # Best trained model

📈 Results

Approach Comparison

We experimented with multiple approaches before achieving optimal results:

Approach	R² Score	MAE (months)	Notes
EfficientNet-B4 (frozen)	0.70	33	Insufficient adaptation
Xception + CLAHE	-0.01	N/A	Preprocessing broke transfer learning
Xception + ROI crop	0.68	28	Lost contextual information
Our Final Model	0.9169	9.04	All layers trainable ✓

Key Insights

1. Architecture-specific preprocessing is critical - Using xception.preprocess_input() was essential
2. Training all layers from start outperformed gradual unfreezing approach
3. Contrast enhancement (CLAHE) breaks transfer learning from ImageNet weights
4. Simple regression heads (10 units) work well with strong base models

Visualizations

Predicted vs Actual Age

• Strong linear correlation (R²=0.9169)
• Most predictions within ±12 months clinical threshold

Residual Plot

• Errors centered around zero
• Slight heteroscedasticity at age extremes

Grad-CAM Attention

• Model correctly focuses on carpal bones and epiphyseal growth plates
• Validates medically-correct feature learning

Gender Bias Analysis

• Minimal performance difference between male/female samples
• Fair and unbiased predictions

👥 Team

Pattern Recognition and Machine Learning Course Project

• Amit Anil Kamble - CS23B2034
• Jatin Goyal - CS23B2045
• Sumit Kumar - CS23B2008

Guided By:
Dr. Umarani Jayaraman, Assistant Professor

🎓 Acknowledgments

• RSNA for providing the Pediatric Bone Age Challenge dataset
• Keras Applications for pre-trained Xception weights
• TensorFlow team for the deep learning framework
• Course instructors and peers for valuable feedback

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

📚 References

1. Halabi, S. S., et al. (2019). "The RSNA Pediatric Bone Age Machine Learning Challenge." Radiology.
2. Chollet, F. (2017). "Xception: Deep Learning with Depthwise Separable Convolutions." CVPR.
3. Selvaraju, R. R., et al. (2017). "Grad-CAM: Visual Explanations from Deep Networks." ICCV.

📧 Contact

For questions or collaborations:

• GitHub: @AmitAK1
• Email: cs23b2034@iiitdm.ac.in

---

⭐ If you find this project useful, please consider giving it a star!