Hierarchical deep learning pipeline for robust cervical parameter measurement in radiographs with C7 obscuration.

Abstract

We developed and externally validated a hierarchical deep learning pipeline that automates cervical sagittal measurements, explicitly addressing C7 obscuration on lateral radiographs. The model combines a global keypoint detector with C2/C7 specialists localized via a multilayer perceptron to refine landmarks on high‑resolution patches. Trained on 5604 images and tested internally and on a challenging external cohort enriched for C7 obscuration (82%), it achieved excellent agreement with ground truth. Externally, intraclass correlation coefficients (ICCs) were 0.97 for lordosis (mean absolute error [MAE] 2.6°), >0.99 for C2 slope (MAE 0.8°), and 0.93 for C7 slope (MAE 2.3°), with minimal bias and narrower limits of agreement than a single‑stage baseline. The model showed near-perfect repeatability (ICC > 0.99) and higher artificial intelligence-expert agreement (ICC 0.81-0.84) for C7 slope than inter-expert reliability (ICC 0.67). In failure cases, the pipeline corrected large global model errors (e.g., 10.22°- 0.22°). This robust, coarse‑to‑fine approach advances reliable, generalizable cervical alignment assessment in real‑world conditions.