Spine labeling in MRI via regularized distribution matching

This study investigates an efficient (nearly real-time) two-stage spine labeling algorithm that removes the need for an external training while being applicable to different types of MRI data and acquisition protocols.

Based solely on the image being labeled (i.e., we do not use training data), the first stage aims at detecting potential vertebra candidates following the optimization of a functional containing two terms: (i) a distribution-matching term that encodes contextual information about the vertebrae via a density model learned from a very simple user input, which amounts to a point (mouse click) on a predefined vertebra; and (ii) a regularization constraint, which penalizes isolated candidates in the solution. The second stage removes false positives and identifies all vertebrae and discs by optimizing a geometric constraint, which embeds generic anatomical information on the interconnections between neighboring structures. Based on generic knowledge, our geometric constraint does not require external training.

We performed quantitative evaluations of the algorithm over a data set of 90 mid-sagittal MRI images of the lumbar spine acquired from 45 different subjects. To assess the flexibility of the algorithm, we used both T1- and T2-weighted images for each subject. A total of 990 structures were automatically detected/labeled and compared to ground-truth annotations by an expert. On the T2-weighted data, we obtained an accuracy of 91.6% for the vertebrae and 89.2% for the discs. On the T1-weighted data, we obtained an accuracy of 90.7% for the vertebrae and 88.1% for the discs.

Our algorithm removes the need for external training while being applicable to different types of MRI data and acquisition protocols. Based on the current testing data, a subject-specific model density and generic anatomical information, our method can achieve competitive performances when applied to T1- and T2-weighted MRI images.