Fracture Finder: Computer-aided real-time diagnosis of vertebral fractures in thoracic spine X-rays using YOLOv8 with Weighted Box Fusion and augmentation strategies

Document Type

Article

Publication Date

1-1-2026

Abstract

Accurate and timely detection of vertebral-body fractures on thoracic-spine X-rays is critical, yet subtle injuries are frequently missed in high-pressure radiology workflows. We proposed Fracture-Finder, a single-stage deep-learning system that localizes suspected acute fractures with bounding boxes and returns confidence scores in real-time. Built on the YOLOv8 object-detection backbone, the model processes a typical X-ray in (Formula presented.) ms on a mid-range consumer GPU. The model was trained and validated on 1247 expert-annotated images from the UTMB-1000 dataset. Exploratory data analysis revealed marked class imbalance and distinct trends in bounding-box size and position. To address these issues, we systematically oversampled fracture cases and applied targeted augmentations such as gamma adjustment, CLAHE, mild rotations, and grid distortion, thereby improving robustness while preserving anatomical detail. At inference, we add light test-time augmentation, along with Weighted Box Fusion, to refine box placement without sacrificing speed. On a 20 % hold-out set Fracture Finder attains (Formula presented.) (Formula presented.) 93% precision and 97% recall, correctly flagging 97% of fractures with few false positives. With its high accuracy, modest hardware requirements, and rapid processing on this single-institution dataset, the model appears computationally compatible with future integration into picture archiving and communication system (PACS) workflow. Fracture-Finder offers a reliable automated tool, flagging suspicious cases for expedited review and reducing the risk of overlooked spinal injuries in both routine and emergency contexts.

Publication Source (Journal or Book title)

Iise Transactions on Healthcare Systems Engineering

First Page

1

Last Page

18

This document is currently not available here.

Share

COinS