Bone-Age-Assessment

Overview

This project leverages deep learning to automate bone age estimation in children using hand-wrist radiographs. It is a university project conducted by me, Erica Marras, and my colleague, Daniela Di Labbio, for the course Human Data Analytics.

Dataset

• Source: RSNA Pediatric Bone Age Challenge.
• Size: 12,611 images (balanced gender distribution)
• Preprocessing:
  • CLAHE contrast adjustment.
  • Random augmentation: flips, rotations (±7.2°), brightness (±10%), zoom (0.8x–1.2x).
  • Images resized to 224x224.

Models

CNN

• Baseline model with 3 convolutional layers.
• Outputs single age prediction.

ResNet

• 10 residual blocks with skip connections.
• Global average pooling for robust predictions.

Global-Local ResNet

• Combines global context and local regions of interest.
• Dual-branch for comprehensive feature extraction.

Gender Integration

• One-Hot Encoding: Basic gender representation.
• Gender Embedding: Captures nuanced patterns.

Best Results

Model	MAE (Months)	Accuracy (≤1 Year)
CNN	12.33	57.02%
ResNet	8.01	77.95%
Global-Local ResNet	8.48	73.75%

• Best Model: ResNet with gender embedding.

Technical Details

• Frameworks: Python, TensorFlow.
• Hardware: NVIDIA T4 GPU.
• Loss Function: Mean Squared Error (MSE).