Fractured Scaphoid Detection

Use Faster R-CNN and YOLOv11-OBB to detect the scaphoid fracture location.

Get Started

1. Configuration

   git clone https://github.com/Hlunlun/Fractured-Scaphoid-Detection
   cd Fractured-Scaphoid-Detection
   conda env create --file environment.yml

2. Run
   • Training

     python main.py --train 1

   • Start System

     python main.py

Model

Name	Description	path
ScaphoidDetector	Detects scaphoid bone in X-ray hand images using Faster R-CNN	scaphoid_detector.py
FractureClassifier	Classify scaphoid fractures using VGG16 pre-trained model after detection by ScaphoidDetector	fracture_classifier.py
HandDetector	Detects scaphoid bones and fractures region in X-ray hand image using YOLOv11-OBB	hand_detector.py

Methods

1. ScaphoidDetector + FractureClassifier + HandDetector
   First, use Faster R-CNN to detect the scaphoid bone in the full X-ray hand image. Then, use VGG16 to classify whether there is a fracture. Finally, use YOLOv11-obb to detect the fracture location.

2. HandDetector
   Directly use YOLOv11-obb to detect the scaphoid bone and fracture locations.

ScaphoidDetector + FractureClassifier + HandDetector

Traning Datasets

1. File Structure

   ip_data  
   ├── fracture_detection  
   │   └── annotations         // Fracture locations: rectangle coordinates [[x1, y1], [x2, y2], [x3, y3], [x4, y4]]  
   └── scaphoid_detection  
       ├── annotations         // Scaphoid locations: rectangle coordinates of the upper-left and lower-right corners [x1, y1, x2, y2]  
       └── images              // Hand X-ray images  
   

2. After data preprocessing in dataset.py, all_datas.json and new folders will be created under fracture_detection and scaphoid_detection:

   ├── ip_data
       ├── fracture_detection
       │   ├── annotations
       │   ├── images
       │   └── images_rec
       └── scaphoid_detection
           ├── annotations
           ├── images
           └── images_rec
   ├── all_datas.json
   

   • fracture_detection/
     • images/: Contains the full scaphoid images cropped based on scaphoid locations.
     • images_rec/: Contains the scaphoid images with highlighted fracture locations.

     fracture_detection
     ├── annotations
     ├── images
     └── images_rec
     

   • fracture_detection/images_rec: Stores hand images with the scaphoid region framed.

Training

1. Train ScaogiudDetector

   from scahpoid_detector import ScaphoidDetector
   scaphoid_detector = ScaphoidDetector(args)
   scaphoid_detector.train()

2. Train FractureClassifier

   from fracture_classifier import FractureClassifier
   fracture_classifier = FractureClassifier(args)
   fracture_classifier.train()

3. Train HandDetector

   from hand_detector import HandDetector
   hand_detector = HandDetector(args)
   hand_detector.train()

Detect

Steps 1. Detect Scaphoid

• Use detect() function

  scaphoid_detector.detect(dir_path)

• Detected scaphoid location will be cropped and saved in prediction/scaphoid/

Steps 2. Classify fracture

• Use classify() function

  fracture_classifier.classify(dir_path)

• Fracture scaphoid will be saved in prediction/classifier/

Steps 3. Detect fracture location

• Use detect_fracture() function
• The images with marked fracture locations will be saved in prediction/fracture/

HandDetector

Training Datasets

使用 yolo_anno.py 內的函數來建構 YOLOv11-OBB的資料

1. File Structure

   yolo_config
   ├── data
   ├── datasets
   │   ├── fracture
   │   │   ├── images
   │   │   │   ├── train
   │   │   │   └── val
   │   │   └── labels
   │   │       ├── train
   │   │       └── val
   │   └── hand
   │       ├── images
   │       │   ├── train
   │       │   └── val
   │       └── labels
   │           ├── train
   │           └── val
   └── weights
   

Training

1. Train hand detector

   from hand_detector import HandDetector
   hand_detector = HandDetector(args)
   hand_detector.train()

2. Curve will be in runs/

Results

1. After prediting
   

Detect

Detect the scaphoid bone and fractures in X-ray hand images. Choose the coordinates with the highest confidence and plot the bounding boxes.

System

Load a folder containing the dataset file structure. The system will then begin predicting and save the images with the scaphoid and fracture locations highlighted.

References

• The Detection and Classification of Scaphoid Fractures in Radiograph by Using a Convolutional Neural Network
• Oriented Bounding Box (OBB) Datasets Overview