← Back to Research

📘 Research summary · Journal article

Real-time localization of vertebral fractures on spine X-rays

Fracture Finder: Computer-Aided Real-Time Diagnosis of Vertebral Fractures in Thoracic Spine X-Rays Using YOLOv8 with Weighted Box Fusion and Augmentation Strategies

Islam, M. S., Anik, M. A., Wasi, A. T., Bappy, M. M.

IISE Transactions on Healthcare Systems Engineering, 2026

Fracture Finder pipeline: pre-processing of an imbalanced dataset, YOLOv8s training with on-the-fly augmentation, test-time augmentation with weighted box fusion, and evaluation (mAP, F1, latency).
Fracture Finder pipeline: pre-processing of an imbalanced dataset, YOLOv8s training with on-the-fly augmentation, test-time augmentation with weighted box fusion, and evaluation (mAP, F1, latency).
⚡ TL;DR

Acute vertebral fractures can be missed under time pressure. Fracture Finder is a single-stage YOLOv8 detector that localizes suspected fractures on thoracic spine X-rays with bounding boxes and confidence scores in about 126 ms per image on a mid-range consumer GPU, supporting faster triage of high-risk cases.

Concept animation: single-pass detection on a spine X-ray
suspected fracture Single-stage detector YOLOv8 looks at thewhole image in one pass about 126 ms per X-ray on a mid-range GPU Output bounding box +confidence score flags cases forfaster review

🎯 Problem

Delayed or missed detection of vertebral fractures postpones treatment. Clinical decision support is most useful when it is fast and indicates where an abnormality is located, not only whether one is present.

💡 Approach

Formulate diagnosis as object detection. A single-stage YOLOv8 detector predicts boxes and confidence scores in one pass, targeted oversampling and augmentation address data imbalance, and weighted box fusion consolidates overlapping predictions.

🧩 Method

Method overview

  1. 1

    Annotated data

    1247 expert-annotated thoracic spine radiographs from the UTMB-1000 dataset.

  2. 2

    Exploratory analysis

    Analysis revealed class imbalance and systematic patterns in box size and position.

  3. 3

    Imbalance handling

    Targeted oversampling and augmentation rebalance the training distribution.

  4. 4

    Detection

    YOLOv8 predicts bounding boxes and confidence scores for suspected fractures.

  5. 5

    Inference and fusion

    Test-time augmentation (identity, horizontal flip, scale) and weighted box fusion consolidate predictions, and suspicious cases are flagged for review.

📊 Results

Key findings

~126ms
Inference time per X-ray on a mid-range consumer GPU
1247
Expert-annotated images for training and validation
  • Fracture Finder achieved strong detection performance on a hold-out set.
  • Its low latency makes it suitable as a triage aid that prioritizes suspicious cases for review.

🏭 Significance

Fast, localized detection can shorten time to diagnosis and reduce missed fractures. The same real-time detection design transfers directly to visual inspection in manufacturing.

📄 Citation and links

Islam, M. S., Anik, M. A., Wasi, A. T., Bappy, M. M.. "Fracture Finder: Computer-Aided Real-Time Diagnosis of Vertebral Fractures in Thoracic Spine X-Rays Using YOLOv8 with Weighted Box Fusion and Augmentation Strategies." IISE Transactions on Healthcare Systems Engineering, 2026.

DOI All publications
📚 More research

Other research summaries

📘 Summary Entropy-Aware Model Predictive Control for Transient Quality Mitigation in WAAM Open 📘 Summary Feature-Aware Anisotropic Local Differential Privacy for Graph Learning in Metal AM Open 📘 Summary Beyond local receptive fields: vision transformers for real-time surface defect detection in FDM Open