Automatic Pavlov ratio measurement method based on spinal landmarks identification by a deep-learning model

Background

Cervical canal stenosis is one of the important pathogenic factors of cervical spondylosis. The accuracy of the Pavlov ratio measurement is crucial for the diagnosis and treatment of cervical spinal stenosis. Manual measurement is influenced by observer variability, accompanied by its inefficiency, which affects clinical evaluation.

Purpose

To automatically and accurately measure the Pavlov ratio, we develop a novel deep-learning model by detecting keypoints of cervical spine and measure the Pavlov ratio on plain lateral cervical spine radiographs.

Methods

We developed a two-stage deep-learning model; in the first stage, we employ the YOLOX model as the object detection network to locate the ROIs containing the vertebral bodies and spinous processes; in the second stage, we introduce the high-resolution net (HRNet) as keypoint detection network and a series of deconvolutional networks (DNs) as the heatmap-based regressor. Based on the mentioned combining algorithms, we can rapidly detect the 38 keypoints in plain lateral cervical spine radiographs, and then measure the Pavlov ratio of the cervical spine. Radiographs from Shanghai Changhai Hospital (a total of 874) were split into training and test subsets (787 and 87 radiographs, respectively). One hundred twelve cases from Shanghai Changzheng Hospital and 108 cases from Shanghai Fourth People's Hospital are used as external validation datasets.

Results

Our proposed model successfully achieved the objective of automating the recognition of spinal landmarks with the mean absolute error (MAE)ranged from 0.05 to 0.08, and the symmetric mean absolute percentage error (SMAPE) ranged from 4.54% to 6.43%. The achieved accuracy is comparable to that of seasoned medical professionals and notably surpasses the performance of junior physicians (SMAPE ranged from 8.74% to 26.19%). Furthermore, our model demonstrated excellent accuracy in external validation experiments (SMAPE ranged from 4.40% to 5.95%).

Conclusion

This study presents a novel YOLOX-HRNet-DN model to assist landmarks identification on lateral cervical spine radiographs and demonstrates excellent accuracy in measuring the Pavlov ratio. The proposed method could provide a potential tool for the automatic estimation of the Pavlov ratio to improve the efficiency and accuracy of the treatment workflow.