Information Technology Reference
In-Depth Information
pose a challenge during both segmentation and lytic lesion detection. Disks and
lytic lesions are both low intensity regions, so disks may be mistaken for lesions
and disks may cause volume averaging with the vertebrae. This accounts for 27 %
of false positive detections in the training set. These false positive detections may be
reduced if the intervertebral disks can be automatically identi
ed within the spinal
column.
The 5 mm slice thickness of our dataset posed another challenge to the seg-
mentation of the spine and detection of lesions. This thickness is common for
routine CTs of the chest, abdomen, and/or pelvis. Our system is designed to
nd
unexpected spinal metastases on examinations ordered for other indications. Thick
slices lead to volume averaging, causing parts of the vertebral body to have
intensity similar to surrounding soft tissues. Substantial leakage (segmentation of
undesired structures) could happen when region-based segmentation is applied. We
adopted a multi-pass technique to address this problem. First a high threshold was
applied to get the initial segmentation, then morphological operations and rolling
balls were applied to close the holes and gaps. Thick slices increase
false
positive detections, especially in the vertebral arch, when oblique cuts result in
volume averaging of vertebral cortex and adjacent soft tissues. Finally, the slice
thickness makes lesions more dif
“
normal
”
cult to detect, as most only appear on one slice.
Other low intensity structures that cause false positive detections are the spinal
canal, the basivertebral vein, and the vein
'
s connections to the anterior external
venous plexus. These false positives have characteristic features, especially in terms
of location and shape, which may be used in future CAD systems to recognize and
eliminate them from the CAD potential lesion list that is presented to radiologists
for consideration.
False negatives can be attributed to two main causes. Two of them were due to
failures in segmentation of the pedicle, while the
final false negative was due to a
failure of characterization. Further work could increase the accuracy of segmen-
tation of the pedicle, while a larger training set may help the classi
er distinguish
between true lesions and false positives.
Quantitative metrics for lesion volume and CT attenuation were calculated for
comparison of CAD system and manually performed lesion characterization. The
difference in mean lesion volume and CT attenuation between ground truth and
computer-aided detections was not statistically signi
cant. Thus, the CAD system
was able to quantitatively characterize the detected lesions with good agreement
when compared to the manually segmented data set. Of particular importance,
manual tracing of the margins of each lesion for electronic segmentation, with each
lesion typically extending over multiple axial image slices required hours of radi-
ologists
time for typical cases. The CAD system was able to automatically analyze
each case in less than 2 min on a standard of
'
ce desktop computer.
There were limitations in CAD system design. First, the bounding region for
lesion search was limited to the body and pedicles of the vertebrae. This anatomic
simpli
ed as this work was intended as a preliminary proof of
concept study, and it has been shown previously that the primary loci for (early)
metastatic spread to the vertebrae are in the vertebral body and pedicles, and the
cation was felt justi
