🦴 Bone Age Prediction with Deep Learning

---

📖 Table of Contents

• 🔎 Overview
• 📂 Dataset
• 🧠 Models Implemented
• 🎨 Grad-CAM Explainability
• 📊 Results
• ⚙️ Installation
• Project Structure
• 🔮 Future Work
• 📚 References

---

🔎 Overview

This repository focuses on Bone Age Assessment using advanced Convolutional Neural Networks (CNNs).
Bone age prediction is critical in diagnosing growth disorders in children. The project leverages multiple state-of-the-art architectures such as EfficientNet, ResNet101, and Xception, alongside Grad-CAM visualizations for explainability.

The goal is to:

• Improve accuracy in age estimation from X-ray images.
• Compare different architectures’ performance.
• Provide interpretability via Grad-CAM heatmaps.

---

📂 Dataset

We use the RSNA Pediatric Bone Age dataset, which contains X-ray images of hands annotated with bone ages.

• Source: RSNA Bone Age Challenge
• Preprocessing steps:
  • Rescaling images to a fixed input size.
  • Normalization for stable training.
  • Augmentation (rotation, flipping, scaling) for robustness.

---

🧠 Models Implemented

✅ EfficientNetB0

• Lightweight model, optimized for speed and accuracy.
• Good for baseline comparisons.

✅ EfficientNetB5

• Larger EfficientNet variant with more layers & parameters.
• Higher accuracy at cost of computation.

✅ ResNet101

• Deep residual network with skip connections.
• Handles vanishing gradients effectively.

✅ Xception

• Depthwise separable convolutions for efficiency.
• Strong performance in image classification tasks.

---

🎨 Grad-CAM Explainability

To ensure trustworthy AI in healthcare, we apply Grad-CAM to visualize which regions of X-rays influence predictions.

Example visualization:

---

📊 Results

Model	Classification Model	Regression Model	Notes
EfficientNetB0	87.00%	8.13	Baseline
EfficientNetB5	87.10%	8.37	Best performing EfficientNet
ResNet101	86.95%	9.56	Stable, deep model
Xception	86.58%	6.82	Good trade-off between accuracy and efficiency

---

⚙️ Installation

# Clone this repository
git clone https://github.com/dpavansekhar/AI-for-Predicting-Age-Specific-Human-Skeleton.git
cd AI-for-Predicting-Age-Specific-Human-Skeleton

# Create a virtual environment
python -m venv venv
source venv/bin/activate   # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt 

---

Project Structure

📦 bone-age-prediction
 ┣ 📂 images
 | ┣ Thirteen-short-bones-in-the-hand-and-wrist-observed-in-the-scoring-system-based-on-the_W640.jpg
 ┣ 📂 notebooks
 ┃ ┣ EfficientNetB0.ipynb
 ┃ ┣ EfficientNetB5, ResNet101 & Xception.ipynb
 ┃ ┣ GradCAM.ipynb
 ┃ ┣ RSNA Bone Age Model Testing.ipynb
 ┣ 📂 outputs
 ┃ ┣ 📂 EfficientNetB0
 ┃ ┃ ┣ EffecientnetB0_Reg.keras
 ┃ ┃ ┣ EffecientnetB0.keras
 ┃ ┣ 📂 EfficientNetB5
 ┃ ┃ ┣ EfficientnetB5_reg.keras
 ┃ ┃ ┣ EfficientnetB5.keras
 ┃ ┃ ┣ history_eff_reg.pkl
 ┃ ┃ ┣ history_eff.pkl
 ┃ ┣ 📂 ResNet101
 ┃ ┃ ┣ history_res_reg.pkl
 ┃ ┃ ┣ history_resnet.pkl
 ┃ ┃ ┣ Resnet101_reg.keras
 ┃ ┃ ┣ Resnet101.keras
 ┃ ┣ 📂 Xception
 ┃ ┃ ┣ history_xcp_reg.pkl
 ┃ ┃ ┣ history_xcp.pkl
 ┃ ┃ ┣ Xception_reg.keras
 ┃ ┃ ┣ Xception.keras
 ┣ 📂 research papers
 ┃ ┣ 📚 1.pdf
 ┃ ┣ 📚 2.pdf
 ┃ ┣ 📚 3.pdf
 ┃ ┣ 📚 4.pdf
 ┃ ┣ 📚 5.pdf
 ┃ ┣ 📚 6.pdf
 ┣ 📜 requirements.txt
 ┣ 📜 README.md
 ┗ 🔧 .gitattributes 

---

🔮 Future Work

1. Incorporate Vision Transformers (ViT).
2. Deploy as a Flask/Django Web App for doctors.
3. Explore multi-task learning (gender + bone age).

---

📚 References

1. RSNA Pediatric Bone Age Dataset - Kaggle
2. EfficientNet Paper: Tan & Le, 2019
3. ResNet Paper: He et al., 2016
4. Xception Paper: Chollet, 2017
5. Grad-CAM Paper: Selvaraju et al., 2017

---

Author - Dogga Pavan Sekhar

For any doubts, contact doggapavansekhar@gmail.com