Cross-Modality Vertebrae Localization and Labeling Using Learning-Based Approaches - Spinal Imaging and Image Analysis - page 310

Fig. 7 Examples of spine detection in challenging CT scans. a C-Spine scan with metal artifacts.
b C-Spine scan with motion artifacts. c L-Spine scan with spinal cord disease. d L-Spine scan with
metal implant. e Whole spine scan with metal artifacts. f Whole spine scan with large imaging
noises

7.3 Comparisons

To illustrate the importance of hierarchical learning and the local articulation
model, we also evaluate results from two adapted versions of the proposed
method. In Method1, we take out the hierarchical learning part. Speci cally, ded-

icated detectors are trained for each vertebra and inter-vertebral disc. In Method2,
we take out local articulated model and use the standard PCA-based method to
model the spine geometry (Table 3 ).
In Table 3 , we list the qualitative results of these three methods on 300 MR scout
scans. The proposed method generates “ perfect ” results in more than 97 % cases,

which is signi cantly better than the others. In general, Method2 is better than