Comparison of multistage and single-stage framework for automated landmark localization and radiologic measurement: a case of the C1-2 complex on cervical spine lateral radiographs

Despite recent advances, applying deep learning to X-rays remains challenging due to the need for substantial downsampling and limited dataset sizes. While multistage frameworks have been proposed to promote stable performance, their advantages over single-stage designs remain unclear. This study aims to systematically compare the two designs for analyzing the C1-2 complex on cervical spine lateral radiographs. A total of 1200 dynamic cervical spine X-ray and MRI pairs were collected from a tertiary care institution, and 300 studies from another institution were set aside for external validation. Two deep learning-based frameworks for localizing 8 landmarks and measuring 14 radiologic indices of the C1-2 complex were developed. The first of which is a multistage pipeline that implements coarse-to-fine localization, and the other is a single-stage model that processes full-region images. Five-fold cross-validation and an anatomy-aware multi-term loss were implemented, and performance was compared in terms of localization, segmentation, and measurement accuracy. The multistage framework demonstrated significantly smaller errors for the localization of the basion and C2 landmarks on the test sets (p < 0.001). For vertebral segmentation, the multistage framework achieved better performance (p < 0.001), yielding DSC and IoU of over 0.90 and 0.80, respectively. Radiologic measurements were more precise with the multistage design, exhibiting mean errors of 1–2° for angular measurements and 0.5–1.5 mm for distance measurements. We demonstrate that the multistage framework can consistently achieve superior performance over single-stage counterparts in analyzing cervical spine radiographs. It is expected to provide evidence for architectural design when developing deep learning models for the analysis of musculoskeletal imaging studies. Not applicable.