Information Technology Reference
In-Depth Information
Table 1 Automated vertebral body detection results
Spine regions
Detection results
Image number
Vertebra number
Cervical vertebrae
Correct
6
38
False/miss
2
7
Thoracic vertebrae
Correct
4
53
False/miss
2
3
Lumbar vertebrae
Correct
7
31
False/miss
0
0
Using the detected vertebral number to construct a new graphical model and run
the inference procedure as shown in Algorithm III to detect all vertebrae visible
in the image(s).
3 Experimental Results
We validated the present approach on digitally reconstructed radiographs (DRRs) of
twenty-one cadaver spinal segments, where eight of them were from cervical
region, six of them were from thoracic region and the rest were from lumbar region,
as well as one low-dose X-ray radiography of a scoliotic patient. The DRRs were
constructed from the CT volumes of the associated spinal segments. For each CT
volume, a pair of DRRs consisting of an anterior-posterior (AP) image and a lateral-
medial (LM) image were generated. There are in total 132 vertebrae in the DRRs
(45 cervical vertebrae, 56 thoracic vertebrae, and 31 lumbar vertebrae) and there are
13 vertebrae visible in the low-dose X-ray radiography.
For each pairs of DRRs, we started the detection from the LM image due to the
observation that the vertebral bodies in the LM image were more homogeneous
than those in the AP image. As soon as all the vertebrae were detected from the LM
image, we could apply the same approach to the AP image but with a
xed number
of the vertebrae that is determined from the LM image. For each detection, the user
interactively speci
ed two points as the input to our approach with one picked
around the center of the top vertebra and the other around the center of the bottom
vertebra. Our approach was then used to detect all vertebrae from the input image
pair. The outputs from our approach include the number of vertebrae in the image,
as well as the 3D location and orientation of each vertebra, which are reconstructed
from the associated 2D detection results in both images. Figure
4
shows three
examples of the automated detection of vertebrae in three different anatomical
regions.
The automated vertebral body detection results are presented in Table
1
.
Although our approach had false/miss detection on four pairs of images, the false/
miss vertebra detection rate was low. From the totally 132 vertebrae, our approach
could successfully detect 122 vertebrae, which results in a 92.4 % success rate.
