Information Technology Reference
In-Depth Information
previous stage. Those new feature vectors are ranked by their performance and only
the K top feature vectors are passed to the next stage. The rationale behind this
scheme is that feature vectors with the worst performance in N
1st stage are
unlikely to be in the top group in the stage N after one more feature is added.
Essentially, this method combines the bene
−
t of exhaustive search and forward
stepwise search. Only a limited number of vectors are exhaustively examined in
each stage, and the performances of feature vectors improve from stage to stage.
9 CAD Performance Evaluation
The quality of a CAD system can be characterized by its sensitivity and speci
city
in detecting lesions. We used FROC analysis to evaluate the overall performance of
our CAD systems. The software used for the FROC analysis of the data was the
ROCKIT ROC analysis software subroutine library (ROCKIT; C. E. Metz,
B. A. Herman, C. A. Roe, University of Chicago, Ill;
http://xray.bsd.uchicago.edu/
krl/KRL_ROC/software_index6.htm
), which
fits a bi-normal distribution to data
using a maximum likelihood estimate [
54
]. The operating point was chosen based
on expected sensitivity in the clinical setting.
Analysis for statistical signi
cance of sensitivity value differences between the
training and testing sets was performed via a bivariate chi-square test. CT attenu-
ation values and mean volumes of the manually and computer performed lesion
segmentations were compared via mixed model ANOVA analysis taking into
consideration of multiple lesions in one patient. P < 0.05 was considered statisti-
cally signi
cant. XLSTAT (
www.xlstat.com
)
was used for data analysis.
10 Data Sets
The CAD system was tested on two cohorts of patients. The
first cohort has pri-
marily lytic metastasis and the second cohort has primarily sclerotic metastasis.
10.1 Lytic Metastasis Cohort
The study group consisted of 50 patients (30 men and 20 women, age range 18
82,
mean age 54.8), divided into training and test cases (29 and 21 patients, respectively).
Patients carried the diagnoses of melanoma, renal cell carcinoma, prostate cancer,
lung cancer, lymphoma, breast cancer, pheochromocytoma, or other disorders (19, 10,
4, 4, 2, 2, 2, and 7 patients, respectively. Each patient was scanned with either a
4-detector (Lightspeed QX/I, GE Healthcare, 20 patients), 8-detector (Lightspeed
Ultra, GE Healthcare, 29 patients) or 16-detector (Mx8000 IDT, Phillips, 1 patient)
CT scanner. Images were obtained at 5 mm slice thickness. Data sets consisted of an
-
