Biomedical Engineering Reference
In-Depth Information
Experiments:
The authors performed 3D segmentation of echocardio-
graphic real-time three dimensional (RT3D) ultrasound data for extraction of
ventricular volumes. Prior to segmentation, a clinician roughly defined endocar-
dial contours on a limited set of short-axis views. The polygonal surface defined
by this manual tracing was then used to initialize the level set segmentation
process. Deformation of the moving front required about 40s for a single vol-
ume on a PC and initial manual tracing less than 1 min. The authors performed
three experiments for assessment of the method accuracy and robustness to the
initialization:
1.
In the first experiment the authors performed multiple segmentation of the
same RT3D volume using manual initialization from six different users. Compar-
ing volume measurements to precise manual tracing they report a mean square
error of 3.8% and a maximum error of 4.38%. They concluded from these results
that the segmentation is rather sensitive to the manual initialization.
2.
In the second experiment, the authors segmented invitro phantom data
of 18 balloons filled with water and immersed in a tank of water. To take into
account the sensitivity of the segmentation technique to the initialization, two
operators performed two separate tracings and the segmentation was run with
these different initializations. A linear regression coefficient of 0.99 was reported
between true volume values and measurements from 70 level set segmentations.
The standard error of estimate was equal to 9.35 ml, the average error of mea-
surement was
−
2.63 ml (std 10.81 ml). Intraobserver variability was estimated
for each operator as: 1.66% and
−
1.36%. Interobserver variability was estimated
as 1.63%. This experiment reported a maximum error of measurement of 40 ml
for large volumes (above 200 ml). Errors of measurements decreased signifi-
cantly with balloons' true volumes suggesting that ventricular volumes can be
accurately measured with this technique for physiological volume ranges.
3.
An invivo study was also performed on about 18 RT3D cases (exact num-
ber not specified in the paper) using again two different manual tracings from
two operators. The ventricular volume range for this study was [151-467 ml]
which suggest that it included dilated cardiomyopathy patients with enlarged
ventricular cavity. Volume measurements were compared to measurement from
manual tracing on MRI. Linear regression was performed with a correlation
coefficient of 0.97. The standard error of estimate was 20.13 ml, average error
was
−
15.58 ml (std 20.55 ml). Intraobserver variability was 0.16% and
−
2.04%,
and inter-observer variability was
−
2.16%. Ejection fraction measurements were
