Hand Bone Fracture Detection Using YOLOv8

Comparative Analysis of YOLOv8s and YOLOv8m with Oversampling Strategy for Hand Bone Fracture Detection on X-Ray Images Using the FracAtlas Dataset

This project is developed for a Sinta 3 journal submission. The goal is to systematically compare YOLOv8s (Small) and YOLOv8m (Medium) for detecting hand bone fractures on X-ray images from the FracAtlas dataset, using a medically-optimized preprocessing pipeline with oversampling to handle class imbalance.

---

Research Overview

Property	Value
Task	Object Detection (Fracture Localization)
Dataset	FracAtlas — Hand subset only
Models Compared	YOLOv8s (Small) vs. YOLOv8m (Medium)
Class Imbalance Handling	Oversampling on training set
Framework	Ultralytics YOLOv8
Evaluation	mAP@50, Recall, Precision, F1, FPS, Confusion Matrix

---

Dataset

Source: FracAtlas (Hand Subset Only)

Category	Count	Percentage
Hand Fractured	~267	~17.4%
Hand Non-fractured	~1,271	~82.6%
Total Hand Images	~1,538	100%

Original FracAtlas contains 4,083 X-rays across multiple body parts (Hand, Leg, Hip, Shoulder, Mixed). This project uses hand images only to eliminate confounding features from other anatomical regions.

After Oversampling (Training Set Only)

Category	Count	Percentage
Fractured (train)	920	54.4%
Non-fractured (train)	770	45.6%
Total training images	1,690	—
Val images	362	—
Test images	362	—
Total dataset	2,414	—

Validation and Test sets use the original unmodified distribution to ensure unbiased evaluation.

---

Models

Property	YOLOv8s (Small)	YOLOv8m (Medium)
Parameters	~11.2M	~25.9M
GFLOPs (640px)	28.6	79.1
Pretrained on	COCO (80 classes)	COCO (80 classes)
Fine-tuned for	1 class (fracture)	1 class (fracture)

---

Results

Accuracy Metrics (Val Set — Best Epoch)

Metric	YOLOv8s	YOLOv8m
mAP@50	0.847	0.838
mAP@50-95	0.444	0.429
Recall	0.780	0.726
Precision	0.891	0.912
Best Epoch	100	98

Inference Speed (GPU: RTX 3050 4GB)

Model	FPS	Latency (ms)
YOLOv8s	TBD	TBD
YOLOv8m	TBD	TBD

Results to be filled after running Cell 8 in the training notebooks.

---

Implementation Pipeline

Notebook Sequence

01b_hand_dataset.ipynb
  → Filter hand-only images from FracAtlas
  → Apply stratified split (70/15/15)
  → Apply oversampling on training set
  → Output: yolo_dataset_hand_oversampled/

02a_training_yolov8s_oversampled.ipynb
  → Train YOLOv8s on oversampled dataset
  → Benchmark FPS (Cell 8)
  → Evaluate on test set + confusion matrix (Cell 9)
  → Qualitative visualization (Cell 10)
  → Output: runs-old-v8s-oversampled/

02a_training_yolov8m_oversampled.ipynb
  → Train YOLOv8m on oversampled dataset
  → Benchmark FPS (Cell 8)
  → Evaluate on test set + confusion matrix (Cell 9)
  → Qualitative visualization (Cell 10)
  → Output: runs-old-v8m-oversampled/

Running Order

1. Run 01b_hand_dataset.ipynb to prepare dataset (if not done yet)
2. Run 02a_training_yolov8s_oversampled.ipynb — training already complete, run Cells 8–10 for journal outputs
3. Run 02a_training_yolov8m_oversampled.ipynb — training already complete, run Cells 8–10 for journal outputs

---

Project Structure

FracAtlas/
│
├── images/                              # [git-ignored] Original FracAtlas images
│   ├── Fractured/
│   └── Non_fractured/
│
├── Annotations/YOLO/                    # YOLO format label files (.txt)
│
├── yolo_dataset_hand_oversampled/       # [git-ignored] Prepared dataset
│   ├── train/images/ & labels/
│   ├── val/images/   & labels/
│   └── test/images/  & labels/
│
├── runs-old-v8s-oversampled/            # [git-ignored] YOLOv8s training outputs
│   └── detect/fracatlas_yolov8s_adamw/
│       ├── weights/best.pt
│       ├── weights/last.pt
│       ├── results.csv
│       └── confusion_matrix.png
│
├── runs-old-v8m-oversampled/            # [git-ignored] YOLOv8m training outputs
│   └── detect/fracatlas_yolov8m_adamw/
│       ├── weights/best.pt
│       ├── weights/last.pt
│       ├── results.csv
│       └── confusion_matrix.png
│
├── notebooks/
│   ├── 01b_hand_dataset.ipynb           # Data preparation + oversampling
│   ├── 02a_training_yolov8s_oversampled.ipynb   # YOLOv8s training + eval
│   └── 02a_training_yolov8m_oversampled.ipynb   # YOLOv8m training + eval
│
├── fracatlas_hand_oversampled.yaml      # YOLO dataset config (oversampled)
├── fracatlas_hand.yaml                  # YOLO dataset config (original)
├── paper_draft.md                       # Research paper draft
├── dataset.csv                          # Full dataset metadata
├── requirements.txt
├── .gitignore
└── README.md

---

Training Configuration

Hyperparameters (Both Models)

Parameter	Value	Notes
optimizer	AdamW	More stable than SGD for small medical datasets
lr0	0.001	Lower than default (0.01) to avoid overshooting
lrf	0.01	Final LR ratio with cosine scheduler
epochs	100	With early stopping (patience=20)
imgsz	640	Balanced detail vs. VRAM constraint (4GB)
batch	4	Limited by VRAM
cos_lr	True	Smooth cosine decay
warmup_epochs	5	Extended warmup for stability
amp	True	Automatic Mixed Precision

Augmentation Strategy (X-Ray Optimized)

Parameter	Value	Reason
degrees	10.0°	Simulate patient positioning variation
fliplr	0.5	Safe for symmetric hand anatomy
flipud	0.0	Disabled — proximal/distal bone orientation matters
shear	0.0	Disabled — anatomical distortion
perspective	0.0	Disabled — X-rays are orthogonal projections
mosaic	0.5	Reduced from 1.0 — avoids medical image artifacts
mixup	0.1	Light regularization
copy_paste	0.0	Disabled — invalid for medical domain
hsv_h / hsv_s	0.0	Disabled — X-rays are grayscale
hsv_v	0.2	Simulate X-ray exposure variation

---

Journal Outputs Checklist

The following outputs are required for journal submission (Sinta 3):

Output	Source	Status
mAP@50, mAP@50-95, Recall, Precision	results.csv	✅ Done
Training loss curves (box/cls/dfl)	results.png	✅ Done
Confusion matrix (val set)	confusion_matrix.png	✅ Done (auto-generated by YOLO)
Inference FPS	Notebook Cell 8	⏳ Run Cell 8
Confusion matrix (test set)	Notebook Cell 9	⏳ Run Cell 9
Qualitative prediction samples	Notebook Cell 10	⏳ Run Cell 10

---

Environment

Component	Version
Python	3.10.19
PyTorch	2.7.1+cu118
Ultralytics	8.4.63
CUDA	11.8
GPU	NVIDIA RTX 3050 Laptop (4GB VRAM)
OS	Windows 11

Setup

# Create and activate conda environment
conda create -n yolo python=3.10
conda activate yolo

# Install dependencies
pip install -r requirements.txt

# Or install manually
pip install ultralytics torch torchvision --index-url https://download.pytorch.org/whl/cu118

---

Medical Imaging Considerations

Why Recall is Prioritized

In fracture detection, a False Negative (missed fracture) is far more dangerous than a False Positive (unnecessary follow-up). A model with high Precision but low Recall is clinically unacceptable. This is why:

• Oversampling is applied to prevent the model from biasing toward "non-fracture"
• Recall is the primary clinical metric in Results and Discussion

Confidence Threshold Guidelines

Scenario	Threshold	Goal
Screening (IGD)	0.10 – 0.20	Maximize recall, minimize missed fractures
Diagnostic support	0.25 – 0.40	Balanced precision and recall
Clinical decision	≥ 0.50	High precision for direct clinical use

---

References

• FracAtlas dataset: Shadmand et al. (2023). FracAtlas: A Dataset for Fracture Classification, Localization and Segmentation. Scientific Data.
• Ultralytics YOLOv8: Jocher, G. et al. (2023). Ultralytics YOLOv8. GitHub.
• Class imbalance: Buda, M., Maki, A., & Mazurowski, M.A. (2018). A systematic study of the class imbalance problem in convolutional neural networks. Neural Networks, 106, 249–259.

---

License

For research and educational purposes only. Ensure compliance with institutional data regulations when using patient imaging data.