Automated Vertebral Segmentation and Measurement of Vertebral Compression Ratio Based on Deep Learning in X-Ray Images

Abstract

Vertebral compression fracture is a deformity of vertebral bodies found on lateral spine images. To diagnose vertebral compression fracture, accurate measurement of vertebral compression ratio is required. Therefore, rapid and accurate segmentation of vertebra is important for measuring the vertebral compression ratio. In this study, we used 339 data of lateral thoracic and lumbar vertebra images for training and testing a deep learning model for segmentation. The result of segmentation by the model was compared with the manual measurement, which is performed by a specialist. As a result, the average sensitivity of the dataset was 0.937, specificity was 0.995, accuracy was 0.992, and dice similarity coefficient was 0.929, area under the curve of receiver operating characteristic curve was 0.987, and the precision recall curve was 0.916. The result of correlation analysis shows no statistical difference between the manually measured vertebral compression ratio and the vertebral compression ratio using the data segmented by the model in which the correlation coefficient was 0.929. In addition, the Bland-Altman plot shows good equivalence in which VCR values are in the area within average ± 1.96. In conclusion, vertebra segmentation based on deep learning is expected to be helpful for the measurement of vertebral compression ratio.

Keywords: Deep learning; Segmentation; Vertebral compression fracture; Vertebral compression ratio.

Fig. 1

Image processing and zero paddings images of the vertebra. a Original image, b CLAHE and Gaussian filter, c Padding image

Fig. 2

Multi dilated recurrent residual U-Net architecture

Fig. 3

Illustrations of a dilated convolution

Fig. 4

Multi dilated residual block

Fig. 5

Recurrent residual block

Fig. 6

Method for calculating vertebral compression ratio

Fig. 7

Images comparing between ground truth images and predict images. a Original images, b ground truth images, c U-Net, d SegNet, e R2U-Net, f E-Net, g MDR2U-Net

Fig. 8

Receiver operating characteristic curve and precision recall curve

Fig. 9

Comparison of manual VCR and measured VCR with predicted images. a Scatter plot, b Bland–Altman plot