Vertebrae localization, segmentation and identification using a graph optimization and an anatomic consistency cycle

Di Meng, Edmond Boyer, Sergi Pujades

Abstract: Highlights•Different from the existing sequential pipelines, a cycling scheme is proposed to recursively localize, segment and identify vertebrae with the anatomical consistency enforced.•The statistical shape priors is leveraged with deep neural networks to detect the non-standard cases.•We propose a graphical model to enforce the anatomical ordering over individual predictions and detect the transitional vertebrae — the presence of T13, L6 and the absence of T12.•The method either output an anatomically coherent result or report the inconsistency region for the doctors to diagnose.