Information Technology Reference
In-Depth Information
7 Results
7.1 Data Descriptions
We evaluated our method on both CT and MR datasets. Our CT data includes 189
randomly selected CT cases with partial or whole spine coverage. A wide range of
imaging parameters are used for these dataset. For example, slice thickness ranges
from 0.75 to 20 mm and different reconstruction kernel (B30s, B31f, B35f, B60f,
H20f, H31s, H70h) are used. Our MR data includes 300 T1-weighted 3D MR scout
scans. These scout scans have relatively low but isotropic resolution 1.7 mm and
large
fields of view. It is not designed for diagnostic purpose but for vertebrae and
disc detections. Our MR data also covers partial or whole spines. These datasets
come from different clinical sites and were generated by different types of Siemens
MR Scanners (Avanto 1.5T, Verio 3T, Skyra 3T, etc.).
7.2 Evaluations
The automatic spine detection results are shown to experienced radiologists and
rated as
“
perfect
”
(no manual editing required),
“
acceptable
”
(minor manual editing
required) and
(major manual editing required).
As shown in Table
2
, our method reaches
“
rejected
”
cases. It
is worth noting that our testing set includes CT/MR scans with severe diseases or
imaging artifacts. However, our method can still detect vertebrae and discs
robustly. Figures
7
and
8
show the results of our method on challenging CT and
MR scans. As shown in these cases, our method is robust to different kinds of
imaging artifacts (Figs.
7
a, b and
8
c, g), large imaging noises (Fig.
7
f), metal
implants (Figs.
7
d, e and
8
d, f), severe scoliosis (Fig.
8
a), pathologies (Figs.
7
c and
8
e) congenital abnormality (Fig.
8
b) and scans that have anchor vertebrae out of
“
perfect
”
detection in 95
þ %
field of view (Fig.
8
h).
We also conduct quantitative evaluation on 355 discs and 340 vertebrae
from 15 WholeSpine MR scans. The average translation errors of discs and ver-
tebrae are 1.91 and 3.07 mm. The average rotation errors of discs is 2.33
°
.
Table 2 Evaluations of spine detections in CT and MR scans
Number of cases
Perfect
Acceptable
Reject
CT
189
180 (95.2 %)
6 (3.2 %)
3 (1.6 %)
MR
300
293 (97.7 %)
4 (1.3 %)
3 (1.0 %)
