Biomedical Engineering Reference
In-Depth Information
volumes to the thresholded membership values of the corresponding voxels
(with reference to WM, GM, and CSF) above 0.5 from the database ground truth
data. The authors computed TP, FP rates and total volume ratios (VR) to com-
pare segmented volumes and thresholded ground truth volumes. Results with
[TP FP VR] were: [92.3% 2.0% 96.3%] for (WM
+
GM), [92.8% 6.0% 103.2%] for
cortical GM and [92.4% 3.3% 98.1%] for WM showing good performance of the
algorithm in isolating brain tissue and segmenting the cortex.
2.
A second experiment was performed using 20 T1-weighted spoiled gradi-
ent MRIs of normal brain subject from the Internet Brain Segmentation Reposi-
tory (IBSR) of the Center for Morphometric Analysis at the Massachusetts Gen-
eral Hospital [101]. Cases were acquired with two different scanners and all
registered to a standard 3D brain coordinate system. Manual segmentation from
medical experts was available. An overlap metric was defined by the IBSR to
evaluate the performance of automatic segmentation methods, measuring the
ratio TP/(1
+
FP). Such ratio ranges from 0 for no agreement to 1 for perfect
agreement with the manual segmentation considered as the ground truth. MRI
data was interpolated from 3 mm thick coronal slices (as provided) to 1 mm
thickness, achieving isotropic voxels. The GM overlap metric on the whole brain
was 0.657. It outperformed other reported segmentation performances on the
same data sets from [102] with overlap ranking from 0.47 to 0.56. Overlap metric
specifically computed on the cerebral cortex (excluding brain stem and cere-
bellum) was further improved to 0.701. The authors pointed out that if applied
to the phantom data, the overlap metric was 0.875 that compared to the manual
segmentation variability of 0.876 reported in the IBSR database.
3.
A third experiment was performed for a study of the frontal lobe anatomy
on 7 high-resolution MRI data sets acquired with SPGR with isotropic voxel
size of 1.2 mm
3
. The patient population of this study included young autistic
and control subjects for comparison of frontal lobe volumes. The MRI volumes
were preprocessed for inhomogeneity correction. Segmentation was performed
with coupled level set functions and the frontal lobe was manually isolated with
anatomical landmarks. Segmentation accuracy was compared to expert manual
tracing. TP and FP on entire frontal lobe averaged 94.1% and 2.1%. TP and FP
for cortical GM on frontal lobe averaged 86.7% and 20.8%. The authors further
evaluated the reliability statistics on the volume measurements obtained on
the segmented frontal lobe volumes using the method proposed by Schultz and
Chakraborty [103]. The agreement between the expert tracing and the level set
