A multi-stage approach for high-precision measurement of cervical curvature in X-ray images

Abstract

As people’s lifestyles change, cervical spondylosis affects people more often. Early diagnosis is thus crucial in the management of cervical spondylosis. Cervical curvature abnormalities may serve as an objective diagnostic standard in radiographs, commonly assessed by the Cobb angle measurement in clinical practice. The method involves the clinician marking four corners of each vertebra and then drawing lines to calculate the Cobb angle using specific rules. The whole process takes a long time and mainly relies on subjective experience. In this paper, we propose an automatic multi-stage method for measuring cervical spine Cobb angle, which is referred to as MSM-CSCA that can perform the precise measurement of cervical curvature. In Stage 1, the region of interest (ROI) of the cervical spine is carefully cropped automatically. In Stage 2, an improved CA-UNet is used to achieve detailed segmentation of the cervical spine region for the ROI region, which extracts features from the edges of the vertebrae and integrates them with the original ROI to highlight the vertebrae edges more distinctly. At Stage 3, the integrated features are input into a dual-branch landmark detection network, with segmentation and detection branches. Within this stage, the segmentation branch convolutional results are fed back into the detection branch after feature augmentation to provide a more accurate detection area. Finally, the accurately identified cervical landmarks from this network are used to calculate the Cobb angle. Experimental results show that the proposed method achieves SMAPE 15.05% and MAE 3.19 on X-ray images of the cervical spine, significantly better than the other benchmark methods. The proposed method can provide clinicians with an efficient, accurate, and reliable way for the estimation of cervical curvature.