Information Technology Reference
In-Depth Information
19. outsideIntensityDev: standard deviation of intensity outside the detection
20. innerOuterContrast: difference between mean Intensity and outside Intensity
21. neighborIntensity: intensity of neighboring blobs from watershed algorithm.
22. distToBoundary: minimum distance from center of the detection to the border
of vertebra
23. relCoordx: relative x coordinate to the center of spinal canal
24. relCoordy: relative y coordinate to the center of spinal canal
25. onPedicle: whether the detection is inside pedicle region
26. outerBorderRatio: ratio of total borders that are boundary of vertebra
27. corticalBorderRatio: ratio of total borders that are cortical shell of vertebra
28. cordBorderRatio: ratio of total borders that are spinal canal.
filters are then applied to reduce the number of detections and relieve the
burden of the classi
Feature
filters are designed based on obser-
vation of typical bone metastases so that all true detections remain. The set of
er in the next step. The
filters
currently in use are,
￿
Shape
filter: aspectRatio10 < 3.5, to eliminate elongated detections
filter: surfaceArea > 0.05 cm 2 , to eliminate small detections
￿
Size
￿
Density
filter: innerOuterContrast > 150, to eliminate less prominent detections.
filters are loosely set and can reduce more than 50 % of false detections
without impairing sensitivity.
These
8 Machine Learning and Classification
Machine learning techniques transfer expert
'
s knowledge into computer algorithms.
In a CAD system, a classi
er is a mathematical model that determines whether a
detection is a true or false
finding. The classi
er in a CAD system is usually a
supervised learning system, i.e., the classi
er is trained using annotated data by
experts. Well-known classi
ers such as neural networks (NN) [ 47 ] and support
vector machines (SVM) [ 48 ] have been widely used in CAD systems [ 49 ].
SVM is a relatively new technique for data classi
cation. It uses hyperplanes in a
high dimensional feature space to separate data into different classes. SVM is
trained with a learning system derived from statistical learning theory, and is
generalizable to unknown data. In the training phase, detections are given a class
label (lesion, non-lesion) to form feature-class pairs (x, y). Given a training set of
S detections (x 1 , y 1 ), (x 2 , y 2 ),
p
,(x s , y s ), for p-dimensional feature space x i 2<
and y i
, a hyperplane can be optimized to separate the two groups of
data (true and false).
f
1
;
1
g
w T
f
ð
x
Þ ¼
x
Þþ
b
¼
0
ð 23 Þ
Search WWH ::




Custom Search