Information Technology Reference
In-Depth Information
Fig. 1 Challenges of vertebrae labeling in clinical cases. a AMR scan of a scoliosis patient. b ACT
scan of a scoliosis patient. c A MR scan with folding artifact. d A CT scan with metal implant
hierarchical strategy to learn detectors for anchor vertebrae, bundle vertebrae and
inter-vertebral discs, respectively. Speci
cally, different learning strategies are
designed to learn anchor, bundle and disc detectors considering different levels of
“
of these anatomies. At run-time, these detectors are invoked in a
hierarchial way. Second, a local articulation model is designed to describe spine
geometries. It is employed to fuse the responses from hierarchical detectors. As the
local articulation model satis
distinctiveness
”
es the intrinsic geometric characteristics of both
health and disease spines, it is able to propagate information from different detectors
in a way that is robust to abnormal spine geometry. With the hallmarks of hier-
archical learning and local articulated model, our method becomes highly robust
to severe imaging artifacts and spine diseases.
3 Problem Statement
Notations: Human spine usually consists of 24 articulated vertebrae, which can be
grouped as cervical (C
1
-
L
5
) sections.
These 24 vertebrae plus the fused sacral vertebra (S
1
) are the targets of spine
labeling in our study.
We de
C
7
),
thoracic (T1-T12)
1
-
T
12
) and lumbar (L
1
-
ne vertebrae and inter-vertebral discs as V
¼
f
v
i
j
i
¼
1
...
N
g
and
D
¼
f
d
i
j
i
¼
1
...
N
1
g
, where v
i
is the i-th vertebra and d
i
is the inter-vertebral disc
between the i-th and i + 1-th vertebra. Here, vi
i
2
R