Application of deep learning in analysing morphological parameters of cervical computed tomography scans

Objectives The process of measuring morphological parameters of the cervical spine through computed tomography (CT) is repetitive and time-consuming. Deep learning (DL) improves efficiency and consistency. In this study, we built a series of DL-based segmentation algorithms to automatically measure five key parameters for the evaluation of cervical spondylotic myelopathy. Subsequently, we compared the performance of our algorithm with that of physicians to assess its accuracy and clinical application value. Methods Cervical spine CT images of 685 patients were divided into a training ( n = 548) and a test set ( n = 137). The training set was used to develop a VB-Net DL model, including a 3D segmentation model of multiple cervical spine subregions and a key point location model on the midsagittal slice of the cervical spine CT. The parameters measured included sagittal vertebral canal diameter (SCD), sagittal vertebral body diameter (VBD), Pavlov’s ratio, transverse vertebral canal diameter (TCD), and osseous spinal canal area (OSCA). Manual measurements were performed by a radiologist and a spinal surgeon. The model’s performance was evaluated using the Mann–Whitney U test, Pearson correlation, mean absolute error, and Bland–Altman plots. Results DL and manual measurements of the Pavlov’s ratio, SCD, VBD, TCD, and OSCA in the test set showed similar accuracy and consistency. The VB-Net model’s Pearson correlation coefficient exceeded 0.8 for most parameters. Conclusions Our VB-Net-based DL approach can effectively approximate the manual measurements of cervical CT morphological parameters in humans, offing physicians an accurate and efficient auxiliary diagnostic tool.