AdvYOLO: Advanced YOLOv8 Application for Bone Pathology Localization and Classification in Wrist X-ray Images

Abstract

In the field of pediatric orthopedics, accurate and timely identification of wrist fractures is vital for effective treatment and recovery. Fractures significantly affect daily activities and can lead to long-term health issues. Especially in areas with limited medical resources, or for doctors with less experience, interpreting X-ray images accurately is challenging. This paper introduces the AdvYOLO algorithm, an enhanced version of YOLOv8, trained on the GRAZPEDWRI-DX dataset to diagnose wrist bone pathologies. The integration of the Dilation-wise Residual (DWR) and Large Separable Kernel Attention (LSKA) modules is critical for improving feature extraction and classification, as they allow more effective processing of complex patterns in X-ray images, leading to more accurate diagnostics. The mAP 50 value of AdvYOLO improved from 63.8% to 68.7%, achieving state-of-the-art performance in wrist detection. Additionally, the paper presents BoneVisionAI, a tool to assist doctors, particularly those with less experience, in accurately interpreting children's wrist X-rays, aiming to reduce diagnostic errors in the healthcare sector.

YOLOv8 architecture

Requirements

• Linux (Ubuntu)
• Python = 3.9
• PyTorch 1.8.2

Environment

  pip install -r requirements.txt

Dataset

Dataset Split

• GRAZPEDWRI-DX Dataset (Download Link)
• Download dataset and put images and annotatation into ./GRAZPEDWRI-DX_dataset/data/images, ./GRAZPEDWRI-DX_dataset/data/labels.

python ./Fracture_Detection/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.

Data Augmentation

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

 python ./Fracture_Detection/imgaug.py

The file content is modified to the desired output file path.

Pre -training model

You can download it through this link ultralytics/ultralytics: NEW - YOLOv8 🚀 in PyTorch > ONNX > OpenVINO > CoreML > TFLite (github.com)

Trained Model

Use gdown to download the trained model from our GitHub:

gdown https://github.com/Happy-Xiang/AdvYOLO/blob/master/weights/best.pt

Validate

python val.py

Experimental Results

Algorithms	DWR	LSKA	Optimizer	Precision	Recall	F1	map50	Params/M	GFLOPs
YOLOv8s	No	No	SGD	0.768	0.579	0.660	0.608	11.1	28.7
YOLOv8s	No	No	Adam	0.678	0.392	0.497	0.431	11.1	28.7
YOLOv8s-D	Yes	No	SGD	0.84	0.539	0.657	0.642	10.8	28.3
YOLOv8s-D	Yes	No	Adam	0.651	0.527	0.582	0.528	10.8	28.3
YOLOv8s-L	No	Yes	SGD	0.775	0.578	0.662	0.598	12.2	29.5
YOLOv8s-L	No	Yes	Adam	0.694	0.491	0.575	0.524	12.2	29.5
AdvYOLO	Yes	Yes	SGD	0.795	0.584	0.673	0.687	11.9	29.1
AdvYOLO	Yes	Yes	Adam	0.775	0.475	0.589	0.492	11.9	29.1

Examples of pediatric wrist fracture detection on X-ray images. (a) manually labeled images, (b) predicted images.

Application

You can download the corresponding visualization program in this link:

Happy-Xiang/YOLOv8-PySide6-GUI-camera (github.com)