Information Technology Reference
In-Depth Information
collected. Then radiologist
is diagnostic knowledge is applied to analyze the training
data and to derive detection rules (classi
'
ers). In the testing phase, clinical images
are taken as input. An image segmentation step is performed to extract structures of
interest and limit the search region. Characteristic features of the structures are
computed. Relevant features are then selected and input into a classi
er. With the
aid of the detection rules developed in the training phase, the classi
er distinguishes
true lesions from false lesions, or malign lesions from benign lesions. The pre-
liminary detections are then reported to radiologists for
final decisions.
cation include features traditionally used by
radiologists and high-order features that are not inherently intuitive. Features include
shape features such as circularity, sphericity, compactness, irregularity, elongation,
or density features such as contrast, roughness, and texture attributes. Different
detection tasks need different sets of relevant features. Feature selection techniques,
such as forward stepwise method and genetic algorithm, are applied in the training
phase to choose features for further classi
Characteristic features for classi
ers have been
proposed for different applications, including linear discriminant analysis, Bayesian
methods, arti
cation [
20
]. Several classi
cial neural network, and support vector machine [
21
].
The quality of a CAD system can be characterized by the sensitivity and spec-
i
city of the system detection. Sensitivity refers to the fraction of diseased cases
correctly identi
city
refers to the fraction of disease-free cases correctly identified as negative.
“
Receiver
operating characteristic
ed as positive in the system (true positive fraction, TPF). Speci
”
(ROC) curves are used to describe the relationship between
sensitivity and speci
city. The ROC curves show the true-positive fraction
(TPF = sensitivity) versus the false-positive fraction (FPF = 1
city). In
addition to ROC curves, Free-response ROC (FROC) curves (TPF versus false
positive per case) were proposed to more accurately represent the number of false
positive detections [
22
]. The areas under the ROC and FROC curve are measures of
the quality of a CAD system. There is often a tradeoff between achieving high
speci
speci
−
city and high sensitivity. A successful CAD system should detect as many
true lesions as possible while minimizing the false positive detection rate.
CAD systems have been used to detect lesions in the breast, where the increased
the true positive rate in breast cancer screening and improved the yield of biopsy
recommendations for patients with masses on serial mammograms [
23
,
24
]. CAD
has been shown to improve radiologists
performance detecting lung nodules on
chest radiography and CT [
25
,
26
] and to increase sensitivity for detecting polyps
on CT colonography [
27
]. In spine imaging, CAD systems had been developed to
detect spine abnormality and disease such as lytic lesions [
28
], sclerotic lesion [
29
],
fractures [
30
], degenerative disease [
31
], syndesmophyte [
32
] and epidural masses
[
33
]. Several commercial systems in mammography, chest CT and CT colonog-
raphy have already received FDA approval for clinical use.
CAD involves all aspects of medical image processing techniques. For instance,
image segmentation and registration are necessary for feature computation, and
image visualization and measurement are essential to present the results to clinicians.
'
