Development of a YOLOv3-Based Model for Automated Detection of Thoracic Ossification of the Posterior Longitudinal Ligament and the Ligamentum Flavum on Plain Radiographs

TL;DR

This paper introduces a deep learning model based on YOLOv3 to detect spinal ossifications in thoracic X-rays, aiming to improve early diagnosis and screening.

Contribution

A novel YOLOv3-based model is developed for automated detection of OPLL and OLF on thoracic radiographs, achieving high diagnostic accuracy.

Findings

The model achieved 80.6% accuracy and 92.6% recall in detecting OPLL and OLF.

It outperformed spine surgeons in accuracy and recall for combined OPLL and OLF cases.

Detection accuracy was 86.3% for combined OPLL and OLF cases.

Abstract

Background/Objectives: This study aims to develop and validate a YOLOv3-based deep learning model for detecting ossification of the posterior longitudinal ligament (OPLL) and ossification of the ligamentum flavum (OLF) on lateral thoracic radiographs, improving early diagnosis and screening accessibility. Methods: A retrospective dataset of 356 lateral thoracic radiographs, including 176 with OPLL or OLF and 180 controls, was annotated by spine surgeons. The YOLOv3 model was trained using data augmentation and evaluated via five-fold cross-validation, with accuracy, precision, recall, and F1-score compared to two spine surgeons. Results: The model achieved 80.6% accuracy, 70.3% precision, 92.6% recall, and 79.9% F1-score, surpassing spine surgeons in accuracy and recall, especially for combined OPLL and OLF cases. Detection accuracy was 81.1% for OPLL, 53.3% for OLF, and 86.3% for combined cases. Conclusions: The YOLOv3-based model provides high accuracy and robust detection of OPLL and OLF on plain radiographs, offering an efficient and accessible screening tool.

Figures