Small Data, Big Impact: A Multi-Locale Bone Fracture Localization on an Extremely Limited Dataset Via Crack-Informed YOLOv9 Variants

Small Data, Big Impact: A Multi-Locale Bone Fracture Localization on an Extremely Limited Dataset Via Crack-Informed YOLOv9 Variants

Abstract

Automated wrist fracture recognition has become a crucial research area due to the challenge of accurate X-ray interpretation in clinical settings without specialized expertise. With the development of neural networks, YOLO models have been extensively applied to fracture detection as computer assisted diagnoses (CAD). However, detection models can struggle when trained on extremely small datasets, which is often the case in medical scenarios. In this study, we utilize an extremely limited fracture dataset and hypothesize that the structural similarities between surface cracks and bone fractures can allow YOLOv9 to transfer knowledge effectively. We show that pre-training YOLOv9 on surface cracks rather than on COCO is how YOLO models are typically pre-trained, and fine-tuning it on the fracture dataset yields substantial performance improvements. We also show that knowledge gained from the surface cracks, requires fewer epochs to converge and minimizes overfitting. We achieved state-of-the-art (SOTA) performance on the recent FracAtlas dataset, surpassing the previously established benchmark. Our approach improved the mean average precision (mAP) score by 3%, precision by 5%, and sensitivity by 6%. The implementation code is publicly available at URL.

YOLOv9 architecture

Citation

If you find our paper useful in your research, please consider citing:

@INPROCEEDINGS{10838409,
    author={Ahmed, Ammar and Manaf, Abdul and Imran, Ali Shariq and Kastrati, Zenun and Daudpota, Sher Muhammad},
    booktitle={2024 International Conference on Frontiers of Information Technology (FIT)}, 
    title={Small Data, Big Impact: A Multi-Locale Bone Fracture Detection on an Extremely Limited Dataset Via Crack-Informed YOLOv9 Variants}, 
    year={2024},
    pages={1-6},
    keywords={YOLO;Wrist;Sensitivity;Neural networks;Bones;Surface cracks;Information technology;X-ray imaging;Overfitting;Biomedical imaging;Fracture detection;Medical x-ray   imaging;Transfer learning;YOLOv9;Object detection},
    doi={10.1109/FIT63703.2024.10838409}}

Requirements

• Linux (Ubuntu)
• Python = 3.9
• Pytorch = 1.13.1
• NVIDIA GPU + CUDA CuDNN

Environment

  pip install -r requirements.txt

Dataset

Dataset Split

• FracAtlas Dataset (Download Link)

• Download dataset and put images and annotatation into ./FracAtlas/data/Split_Authors/images, ./FracAtlas/data/Split_Authors/labels.

    python split.py
  

• The dataset is divided into training, validation, and testing set (70-20-10 %) according to the key patient_id stored in dataset.csv.

• The script then will move the files into the relative folder as it is represented here below.

   FracAtlas
      └── data
        └── Split_Authors
           ├── meta.yaml
           ├── images
           │    ├── train
           │    │    ├── train_img1.png
           │    │    └── ...
           │    ├── valid
           │    │    ├── valid_img1.png
           │    │    └── ...
           │    └── test
           │         ├── test_img1.png
           │         └── ...
           └── labels
                ├── train
                │    ├── train_annotation1.txt
                │    └── ...
                ├── valid
                │    ├── valid_annotation1.txt
                │    └── ...
                └── test
                     ├── test_annotation1.txt
                     └── ...
  

The script will create 3 files: train_data.csv, valid_data.csv, and test_data.csv with the same structure of dataset.csv.

Data Augmentation

• Data augmentation of the training set using the addWeighted function doubles the size of the training set.

  python imgaug.py --input_img /path/to/input/train/ --output_img /path/to/output/train/ --input_label /path/to/input/labels/ --output_label /path/to/output/labels/

For example:

  python imgaug.py --input_img ./GRAZPEDWRI-DX/data/images/train/ --output_img ./GRAZPEDWRI-DX/data/images/train_aug/ --input_label ./GRAZPEDWRI-DX/data/labels/train/ --output_label ./GRAZPEDWRI-DX/data/labels/train_aug/

• The path of the processed file is shown below:

   GRAZPEDWRI-DX_dataset
      └── data
           ├── meta.yaml
           ├── images
           │    ├── train
           │    │    ├── train_img1.png
           │    │    └── ...
           │    ├── train_aug
           │    │    ├── train_aug_img1.png
           │    │    └── ...
           │    ├── valid
           │    │    ├── valid_img1.png
           │    │    └── ...
           │    └── test
           │         ├── test_img1.png
           │         └── ...
           └── labels
                ├── train
                │    ├── train_annotation1.txt
                │    └── ...
                ├── train_aug
                │    ├── train_aug_annotation1.txt
                │    └── ...
                ├── valid
                │    ├── valid_annotation1.txt
                │    └── ...
                └── test
                     ├── test_annotation1.txt
                     └── ...
  

Model

You can get the open source code of YOLOv9 through YOLOv9 official GitHub.

Train

Before training the model, make sure the path to the data in the meta.yaml file is correct.

• meta.yaml

  # patch: /path/to/GRAZPEDWRI-DX/data
  path: 'E:/GRAZPEDWRI-DX/data'
  train: 'images/train_aug'
  val: 'images/valid'
  test: 'images/test'

• Arguments

Key	Value	Description
model	None	path to model file, i.e. yolov9n.pt, yolov9n.yaml
data	None	path to data file, i.e. coco128.yaml
epochs	100	number of epochs to train for
patience	50	epochs to wait for no observable improvement for early stopping of training
batch	16	number of images per batch (-1 for AutoBatch)
imgsz	640	size of input images as integer, i.e. 640, 1024
save	True	save train checkpoints and predict results
device	None	device to run on, i.e. cuda device=0 or device=0,1,2,3 or device=cpu
workers	8	number of worker threads for data loading (per RANK if DDP)
pretrained	True	(bool or str) whether to use a pretrained model (bool) or a model to load weights from (str)
optimizer	'auto'	optimizer to use, choices=SGD, Adam, Adamax, AdamW, NAdam, RAdam, RMSProp, auto
resume	False	resume training from last checkpoint
lr0	0.01	initial learning rate (i.e. SGD=1E-2, Adam=1E-3)
momentum	0.937	SGD momentum/Adam beta1
weight_decay	0.0005	optimizer weight decay 5e-4
val	True	validate/test during training

• Example

  cd Bone_Fracture_Detection_YOLOv9
  yolo train model=yolov9n.pt data=./GRAZPEDWRI-DX/data/meta.yaml epochs=100 batch=16 imgsz=640 save=True device=0 workers=4 pretrained=yolov9n.pt optimizer=SGD lr0=0.01

Trained Model (Weights)

To download the trained model weights with this link Crack-Informed-YOLOv9-E-Weights.

Validate

• Arguments

Key	Value	Description
data	None	path to data file, i.e. coco128.yaml
imgsz	640	size of input images as integer, i.e. 640, 1024
batch	16	number of images per batch (-1 for AutoBatch)
save_json	False	save results to JSON file
save_hybrid	False	save hybrid version of labels (labels + additional predictions)
conf	0.001	object confidence threshold for detection
iou	0.6	intersection over union (IoU) threshold for NMS

• CLI

  yolo val model=/path/to/best.pt data=/path/to/meta.yaml

Experimental Results

• Performance comparison of variants with and without pre-training on surface cracks.

Variant	Default YOLOv9		Crack-Informed YOLOv9
	Test mAP@50	Val mAP@50	Convergence Ep	Train Time
M	0.38	0.44	320	5.68h
C	0.52	0.53	174	2.60h
E	0.45	0.47	174	2.69h
GELAN	0.43	0.52	150	3.23h
GELAN-C	0.37	0.60	174	3.24h
GELAN-E	0.50	0.59	262	3.59h

---

• Evaluation of a single variant YOLOv9-E using different methods.

# of Training Instances	Method	Test mAP@50	Val mAP@50	Test Precision	Test Sensitivity
574	Default	0.45	0.47	0.64	0.50
574	Crack-Informed (Ours)	0.60	0.59	0.89	0.51
574	Pre-trained on COCO	0.45	0.54	0.61	0.44
1148	Brightness + Contrast Augmentation [3]	0.47	0.49	0.61	0.45
5740	Albumentations Augmentation	0.25	0.37	0.45	0.30

Fracture Detection Using YOLOv9