Bone-fracture-detection

Deep learning pipeline for bone fracture detection from X-ray images using YOLOv8, EfficientNet, and UNet. Includes multitask learning, model comparisons, and a Gradio-powered UI for real-time inference.

Dataset

• Source: Kaggle - pkdarabi/bone-fracture-detection
• Includes bounding box annotations for different types of fractures.

Models Used

1. YOLOv8 for Object Detection

Trained for 30 epochs with imgsz=640, batch=32

Best model saved and evaluated on validation + test set

2. Multitask Model (Single)

Backbone: EfficientNetB0 (classification)

Decoder: UNet (segmentation)

Loss:

• Binary: BCEWithLogitsLoss
• Multiclass: CrossEntropyLoss
• Segmentation: DiceLoss

Performance:

• Binary Accuracy (Fracture vs No Fracture): 0.7751
• Multi-Class Accuracy (Fracture Type): 0.7692
• Dice Score: 0.3071
• IoU Score: 0.1987

3. Separate Classification + Segmentation Models

Classification: EfficientNetB1

Segmentation: UNet with ResNet34 encoder

Performance:

• Classification Accuracy: 0.6864
• Segmentation Dice: 0.5105, IoU: 0.5037

Gradio Demo

Here is the final Gradio interface combining YOLOv8 detection, classification, and segmentation: