PaperlayerView exact revision on arXiv

Source-linked AI summary

Fracture Detection in Pediatric Wrist Trauma X-ray Images Using YOLOv8 Algorithm

Rui-Yang Ju, Weiming Cai

arXiv:2304.05071v5cs.CV

TL;DR

The paper addresses pediatric wrist-fracture detection from X-ray images, where interpretation can require specialized expertise. It applies data augmentation to YOLOv8 on the GRAZPEDWRI-DX dataset and develops an application for surgeons. The resulting model achieves a reported mAP 50 of 0.638, exceeding the cited improved YOLOv7 and original YOLOv8 results.

  • Problem

    Pediatric wrist-fracture X-ray interpretation can require time-consuming specialized training and may be difficult where experienced radiologists are unavailable.

  • Method

    The paper applies data augmentation to YOLOv8 for fracture detection using the GRAZPEDWRI-DX pediatric wrist trauma X-ray dataset and develops a supporting application.

  • Results

    mAP 50 of our model is 0.638, higher than 0.634 for improved YOLOv7 and 0.636 for original YOLOv8.

  • Takeaways & Limitations

    The Fracture Detection Using YOLOv8 App aims to assist pediatric surgeons in interpreting X-ray images and reduce the probability of analysis errors.

Abstract

from arXiv · show

Hospital emergency departments frequently receive lots of bone fracture cases, with pediatric wrist trauma fracture accounting for the majority of them. Before pediatric surgeons perform surgery, they need to ask patients how the fracture occurred and analyze the fracture situation by interpreting X-ray images. The interpretation of X-ray images often requires a combination of techniques from radiologists and surgeons, which requires time-consuming specialized training. With the rise of deep learning in the field of computer vision, network models applying for fracture detection has become an important research topic. In this paper, we use data augmentation to improve the model performance of YOLOv8 algorithm (the latest version of You Only Look Once) on a pediatric wrist trauma X-ray dataset (GRAZPEDWRI-DX), which is a public dataset. The experimental results show that our model has reached the state-of-the-art (SOTA) mean average precision (mAP 50). Specifically, mAP 50 of our model is 0.638, which is significantly higher than the 0.634 and 0.636 of the improved YOLOv7 and original YOLOv8 models. To enable surgeons to use our model for fracture detection on pediatric wrist trauma X-ray images, we have designed the application "Fracture Detection Using YOLOv8 App" to assist surgeons in diagnosing fractures, reducing the probability of error analysis, and providing more useful information for surgery.

Introduction

Pediatric wrist fractures create a demanding diagnostic workload, especially where radiologist access is limited. This paper applies YOLOv8 with data augmentation and develops an application intended to assist surgeons with X-ray interpretation.

  • X-ray is the most widely used fracture-imaging device because of its cost-effectiveness.
  • 26% of X-ray images were reportedly misinterpreted, highlighting diagnostic challenges where specialist radiologist access is limited.
  • Deep learning and object-detection models have become active approaches for fracture detection in medical images.
  • The study trains YOLOv8 models on GRAZPEDWRI-DX, then applies data augmentation and compares the resulting models with YOLOv7 variants.
  • The augmented YOLOv8 training method reaches a state-of-the-art mean average precision for fracture detection on GRAZPEDWRI-DX.
  • The proposed application aims to help pediatric surgeons interpret X-ray images without radiologist assistance and reduce analysis errors.

Related Work

Prior work applies deep learning and YOLO-family detectors to fractures across body regions, balancing accuracy with inference speed. The paper positions YOLOv8 pediatric wrist-fracture detection as comparatively underexplored.

  • Deep learning studies have detected fractures in the thigh, arm, hand, and spine using DCFPN, R-CNN, Fast R-CNN, Faster R-CNN, and YOLOv2.
  • YOLOv2 achieved 75.3% average precision on spine-fracture CT images, exceeding Faster R-CNN’s 73.3%, with 27 ms inference per image.
  • YOLO-family models are described as balancing detection accuracy and inference speed for real-time X-ray detection on mobile devices.
  • YOLOv5 has been deployed on mobile phones through the “iDetection” application.
  • Few studies had reported pediatric wrist-fracture detection with YOLOv5, and this paper applies YOLOv8 to that task.

Proposed Method

The proposed method augments pediatric wrist X-ray data by adjusting contrast and brightness, then applies a YOLOv8 architecture with redesigned feature extraction, fusion, and detection components. The model uses multi-scale feature fusion and several computationally efficient modules and training objectives.

  • Data Augmentation: The dataset is divided into training, validation, and test sets, with the training set expanded through data augmentation.The training set grows from 14,204 original images to 28,408 augmented images.
  • Data Augmentation: Contrast and brightness adjustment uses OpenCV’s addWeighted function to enhance bone-anomaly visibility.The same image is used as both inputs, with β set to 0; α controls contrast and γ controls brightness.
  • YOLOv8 Architecture: The YOLOv8 model contains backbone, neck, and head components, with C2f, ConvModule, DarknetBottleneck, and SPPF modules.The backbone uses CSP architecture, while C2f replaces the C3 module used in YOLOv5.
  • Backbone: The backbone reduces computational cost through CSP and fewer stage blocks, while SPPF is adopted to improve inference speed.The model uses 3, 6, 6, and 3 blocks across Stages 1–4.
  • Neck: The neck uses FPN and PAN-based multi-scale feature fusion to combine semantic and location information from different network layers.The adopted FP-PAN removes convolution operations during upsampling to reduce computational cost.
  • Head and Loss: The head separates classification and regression branches, uses an anchor-free design, and trains with task-aligned sample assignment and BCE, DFL, and CIoU losses.CIoU loss additionally considers the aspect-ratio consistency between predicted and ground-truth bounding boxes.

Experiments

The experiments evaluate YOLOv8 on the GRAZPEDWRI-DX dataset using object-detection metrics and compare validation results across classes and optimizers. The reported validation setup includes IoU, precision-recall, and F-score measures.

  • GRAZPEDWRI-DX contains 20,327 pediatric wrist-trauma X-ray images from 6,091 patients, annotated into nine classes with bounding boxes.
  • The dataset was randomly divided into training, validation, and test sets containing 14,204, 4,094, and 2,029 images, respectively.
  • IoU measures the overlap-to-union ratio between generated candidate and ground-truth bounding boxes, with higher values indicating less positional difference.
  • The precision-recall curve plots recall against precision, with each connected point representing a different threshold.
  • The F-score evaluates accuracy by combining precision and recall, while the F1-score is their harmonic mean when β = 1.

Experiment Setup

The experiment setup uses pretrained YOLOv8 models, optimizer comparisons, and resource-conscious training configurations. Fracture detection is evaluated at input image sizes of 640 and 1024.

  • Training uses a pretrained YOLOv8 model from the MS COCO val2017 dataset, with an initial 200-epoch setting and patience of 50.
  • Quantitative fracture-detection comparisons are reported for input image sizes of 640 and 1024, with speed including preprocessing, inference, and post-processing.
  • 100 training epochs were selected after optimizer experiments found the best model epochs mostly concentrated between 50 and 70.
  • The experiments compare YOLOv8 models trained with Adam and SGD optimizers to assess their effects on model performance.
  • Training uses SGD with an initial learning rate of 1×10^-2, weight decay of 5×10^-4, momentum of 0.937, and batch size 16.

Ablation Study

The ablation and comparative experiments evaluate data augmentation, optimizer choices, and model performance for pediatric wrist fracture detection. Data augmentation improves bone-anomaly detection and achieves higher mAP than YOLOv8 and YOLOv7 variants, while some difficult image conditions remain challenging.

  • Data Augmentation: Bone-anomaly mAP 50 increased from 0.11 to 0.169, a 53.6% improvement, after contrast and brightness augmentation.The augmentation improved detection of bone-anomaly and contributed to overall model performance.
  • Optimizer Selection: The SGD optimizer reached the best YOLOv8m performance at epoch 35, compared with epoch 70 for Adam.For augmented training, the dataset expanded to 28,408 images and the authors switched to SGD.
  • Comparative Performance: At input size 640, augmented YOLOv8m mAP 50 rose from 0.621 to 0.629, while YOLOv8l rose from 0.623 to 0.637.The augmented models retained the same parameter counts and FLOPs but had longer CPU inference times.
  • Comparative Performance: The proposed model achieved higher mAP than YOLOv7, YOLOv7-CBAM, and YOLOv7-GAM, demonstrating SOTA performance on the dataset.The comparison is reported in Table 7.
  • Detection Cases: The model detected single fractures well, but metal puncture and dense multiple fractures substantially reduced prediction accuracy.The comparison uses radiologist manual annotations and model predictions.

Application

The authors develop a graphical application that packages the trained fracture-detection model for pediatric wrist trauma X-ray analysis. The application supports image prediction and saving results.

  • Implementation: A GUI application was developed with PySide6, a Qt6-based Python library.PySide6 provides the Qt toolkit used to implement the interface.
  • Application Workflow: The “Fracture Detection Using YOLOv8 App” uses an exported ONNX model and allows users to open images, predict fractures, and save predictions.The selected model is YOLOv8s with an input image size of 1024.

Conclusions and Future Work

The paper applies YOLOv8 with data augmentation to pediatric wrist trauma X-rays and develops an application for fracture detection. The authors provide training materials and plan broader platform and model deployment.

  • Conclusions: YOLOv8 was evaluated on 20,327 pediatric wrist trauma X-ray images from 6,091 patients using randomly divided training, validation, and test sets.The dataset was used to train and evaluate the model.
  • Conclusions: The “Fracture Detection Using YOLOv8 App” aims to assist pediatric surgeons, reduce misclassification probability, and provide information for surgery.The application is currently available for macOS.
  • Future Work: Future plans include deploying different model sizes and extending the application to iOS and Android.These plans are intended to support use on mobile devices.
  • Future Work: The authors provide model-training steps and the trained model through GitHub for possible use as a pre-training model for other body regions.The paper states that this can improve performance when detecting fractures outside the pediatric wrist.

Funding

The authors report that the submitted work received no organizational support.

  • Funding: The authors did not receive support from any organization for the submitted work.

Competing interests

The authors report no financial or proprietary interests in materials discussed in the article.

  • The authors declare no financial or proprietary interests in any material discussed in the article.

Ethics approval

The research reports no involvement of human participants or animals.

  • The study does not involve human participants or animals.
Loading 2304.05071v5…