🦴 Bone Fracture Detection (X-ray Images)

📌 Overview

This project focuses on bone fracture detection using deep learning and computer vision. The dataset consisted of approximately 3,000 X-ray images across various fracture types. During preprocessing, I identified issues with imbalanced classes, missing annotations, and redundant labels, which necessitated dataset cleaning and label remapping before training the YOLOv8 model.

---

⚙️ Dataset Preparation

• Original Dataset: 3,631 training, 348 validation, 169 test images
• Annotation Source: Roboflow + Manual corrections with labelImg
• Problem: Duplicate/imbalanced classes and missing annotations

✅ Steps Taken

1. Annotation Fixing:

   • Added missing bounding boxes manually using labelImg.
   • Verified all labels followed YOLO format (class x_center y_center width height).

2. Class Cleaning:

   • Removed duplicate class (humerus fracture had only 3 samples).
   • Unified the label structure to avoid confusion.

3. Label Remapping:

   • Original data.yaml had 7 classes:

     elbow positive
     fingers positive
     forearm fracture
     humerus fracture
     humerus
     shoulder fracture
     wrist positive
     

   • After cleaning, reduced to 6 valid classes:

     elbow positive
     fingers positive
     forearm fracture
     humerus
     shoulder fracture
     wrist positive
     

4. Class Distribution After Cleaning:

   Class 0: 385 instances (elbow positive)
   Class 1: 606 instances (fingers positive)
   Class 2: 373 instances (forearm fracture)
   Class 3: 362 instances (humerus)
   Class 4: 397 instances (shoulder fracture)
   Class 5: 262 instances (wrist positive)
   

---

🚀 Model Training

• Framework: Ultralytics YOLOv8

• Config:

  nc: 6
  names: ["elbow positive", "fingers positive", "forearm fracture", "humerus", "shoulder fracture", "wrist positive"]

• Techniques used:

  • Patience for early stopping (patience=5 helped reduce overfitting).
  • Cosine learning rate scheduling (cos_lr=True).
  • Data augmentation (flips, rotations, scale).

---

📊 Results

• Detected multiple fractures in unseen X-ray images.
• Successfully identified forearm fracture, shoulder fracture, and finger positive cases.
• Observed class imbalance challenges, but generalization improved after dataset cleaning.

---

🔮 Next Steps

• Expand dataset size (currently ~3K).
• Balance underrepresented classes (e.g., wrist positive).
• Explore 3D X-ray or CT-based approaches for better performance.

---

📂 Project Structure

├── dataset/
│   ├── train/
│   │   ├── images/
│   │   └── labels/
│   ├── val/
│   │   ├── images/
│   │   └── labels/
│   └── test/
│       ├── images/
│       └── labels/
├── runs/
│   └── train/
│       └── bone_fracture_exp/
├── data.yaml
└── README.md

---

🙌 Acknowledgements

• Roboflow for dataset tools
• Ultralytics for YOLOv8