classification of benign and malignant vertebral...
TRANSCRIPT
Lucas Frighetto-Pereira, Guilherme Augusto Metzner,
Paulo Mazzoncini de Azevedo-Marques,
Rangaraj Mandayam Rangayyan,
Marcello Henrique Nogueira-Barbosa
Classification of Benign and Malignant
Vertebral Compression Fractures in Magnetic Resonance Images
Anatomy of the spine
โ MRI sagittal slice
VertebralBody
IntervertebralDisc
VertebralArch
Vertebra
LumbarSpine
UNIVERSIDADE DE
SรO PAULO
Vertebral compressionfractures (VCFs)
โ Partial collapse of vertebral bodies
โ Traumatic VCFs raise no doubt about
their etiology
โBut a recent vertebral collapse withouthistory of significant trauma createsdifficulty in defining the cause of the VCF
UNIVERSIDADE DE
SรO PAULO
Medical diagnosis
โข Young patient with a VCF
โข History of significant acute trauma
โข Usually easy diagnosis
UNIVERSIDADE DE
SรO PAULO
Medical diagnosis
โข Elderly patient with VCF
โข No history of significant acute trauma
โข Diagnosis ?
UNIVERSIDADE DE
SรO PAULO
VCFs without history ofsignificant trauma
โVCFs are the most common type ofosteoporotic fractures
โThe elderly have a high incidence of VCFsrelated to metastatic cancer affecting bone
โMRI is the most commonly used imagingmethod for spinal diseases and earlydetection of fractures
UNIVERSIDADE DE
SรO PAULO
OsteoporoticVCF
MetastaticVCF
T1-WeightedMRI
Clinical classificationof VCFs
โOsteoporotic VCFs
โข classified as Benign VCFs
โMetastatic VCFs
โข classified as Malignant VCFs
UNIVERSIDADE DE
SรO PAULO
Benign VCFs in T1-weighted MRI
โ Partial preservation of normal fatty bone-marrow signal in the vertebral body
โDegeneration of normally rectangularshapes of vertebrae into concave and roughshapes with indentations
โRougher contours than malignant VCFs andnormal vertebrae
UNIVERSIDADE DE
SรO PAULO
Malignant VCFs in T1-weighted MRI
โGlobal reduction of signal intensity or nodular abnormality in the affected vertebral body
โCould result in a posterior convexity without substantial concavities
โMay also cause the contours of vertebrae to be relatively smoothened due to convexity
UNIVERSIDADE DE
SรO PAULO
Normal Benign VCFs Malignant VCFs
Benign vs MalignantVCFs
โBoth tend to create concavities in the vertebral plateaus
โCould cause doubt in the diagnosis
โCorrect classification is critical for planning treatment
UNIVERSIDADE DE
SรO PAULO
Malignant VCF
Benign VCF
Which image has the malignant VCF and which one has the benign VCF?
Objectives
โ Study the characteristics of VCFs in MRI
โ Develop image processing techniques to extractfeatures
โ Classify VCFsNormal
Fractured
Benign
Malignant
UNIVERSIDADE DE
SรO PAULO
Study steps
โ Selection of cases and images
โManual segmentation of vertebral bodies
โ Extraction of features of vertebral bodies
โClassification, validation, and statisticalanalysis
UNIVERSIDADE DE
SรO PAULO
Database
โUniversity Hospital of Ribeirรฃo Preto Medical School โ University of Sรฃo Paulo
โCases and images collected from theRadiology Information System (RIS)
โCases from September 2010 to March 2014
โ Philips 1.5T MRI System โ T1-weighted MRI
UNIVERSIDADE DE
SรO PAULO
Database
โ Lumbar vertebral bodies (L1 to L5)
โMedian sagittal slice
โ TIFF images with 8-bits/pixel
โ 153 exams analyzed, 63 selected
โ 38 women, 25 men
โMean age: 62 years
UNIVERSIDADE DE
SรO PAULO
Database
โ 63 selected exams:
โข At least one VCF per patient
โข The nonfractured vertebral bodies of patientswithout malignant fractures are considered tobe normal
UNIVERSIDADE DE
SรO PAULO
Excluded cases
โVertebral fractures secondary to trauma
โ Infection and avascular necrosis
โ Severe degenerative scoliosis
โ Previous surgeries, radiotherapy, andchemotherapy
UNIVERSIDADE DE
SรO PAULO
Database
L5 L4 L3 L2 L1 Total
Benign VCFs 6 7 9 10 21 53
Malignant VCFs 9 11 10 10 9 49
Normal 26 24 23 22 11 106
Total 41 42 42 42 41 208
UNIVERSIDADE DE
SรO PAULO
Examples ofvertebral bodies
Normal
Benign VCFs
Malignant VCFs
UNIVERSIDADE DE
SรO PAULO
Manual segmentation
MRI exam Vertebral body masks
UNIVERSIDADE DE
SรO PAULO
Software flow chart UNIVERSIDADE DE
SรO PAULO
MRI exam and its maskUNIVERSIDADE DE
SรO PAULO
Detection of the coordinatesof the vertebral bodies
L5
L4
L3
L2
L1
UNIVERSIDADE DE
SรO PAULO
Normalizationof the MR images
. 5x5 disc block
Extraction of blocks ofintervertebral discs
using the mask ROIs as reference
UNIVERSIDADE DE
SรO PAULO
DiscsMean
Normalizationof the MR images
๐๐๐ค๐ผ๐๐ ๐, ๐ =๐๐๐๐๐๐๐๐๐๐๐(๐, ๐)
๐๐๐ ๐๐ ๐๐๐๐
๐๐๐๐ต๐๐๐ ๐, ๐ = 255 ร๐๐๐ค๐ผ๐๐ ๐, ๐ โ min ๐๐๐ค๐ผ๐๐
max ๐๐๐ค๐ผ๐๐ โmin ๐๐๐ค๐ผ๐๐
UNIVERSIDADE DE
SรO PAULO
MRI exam โฉ Mask
โฉ
Processing new image...
UNIVERSIDADE DE
SรO PAULO
Detection of theROIs
UNIVERSIDADE DE
SรO PAULO
ROIs of thevertebral bodies UNIVERSIDADE DE
SรO PAULO
Normal
Benign VCFs
Malignant VCFs
Computation of thefeatures
โ 3 Statistical gray-level features
โ 14 Texture features
โ 10 Shape features
27 Features
UNIVERSIDADE DE
SรO PAULO
Statistical gray-levelfeatures
Coefficient ofvariation
Skewness
Kurtosis
UNIVERSIDADE DE
SรO PAULO
Statistical gray-levelfeatures
โCoefficient of variation (CV )
๐ถ๐ =๐
๐
๐ =1
256
๐=1
256
๐ฅ๐
๐ =1
256
๐=1
256
๐ฅ๐ โ ๐ 2
UNIVERSIDADE DE
SรO PAULO
Statistical gray-levelfeatures
โ Skewness
๐ ๐๐๐ค๐๐๐ ๐ =1
256 ร ๐3
๐=1
256
(๐ฅ๐ โ ๐)3
UNIVERSIDADE DE
SรO PAULO
Statistical gray-level features
โKurtosis
๐๐ข๐๐ก๐๐ ๐๐ =1
256 ร ๐4
๐=1
256
(๐ฅ๐ โ ๐)4
UNIVERSIDADE DE
SรO PAULO
Differences in texturebetween normal and VCFs UNIVERSIDADE DE
SรO PAULO
Malignant VCF
Malignant VCF
Malignant VCF
Malignant VCF
Normal
Benign VCF
Texture features
Gray-level
cooccurrence matrix
14 texture features of
Haralick et al.
UNIVERSIDADE DE
SรO PAULO
Cooccurrence matrix
0 0 1 0 0
0 1 2 1 0
1 2 2 2 1
0 1 2 1 0
0 0 1 0 0
Ex: Image 5x5 pixels,3 gray levels
0 1 2
2 2 4
2 4 2
4 2 2
Distance = 1 pixel
Angle = ยฑ45ยฐ
0 1 2
0
1
2
Number of pixels of intensity 0 thatare at ยฑ45 degrees and distance 1
of pixels of intensity 2
UNIVERSIDADE DE
SรO PAULO
Cooccurrence matrix ๐ ๐, ๐ for
๐ = 0, ๐ = 2
14 texture featuresof Haralick et al.
โAngular second moment (Energy)
๐1 =
๐
๐
๐(๐, ๐) 2
โContrast
๐2 =
๐=0
๐๐โ1
๐2
๐=1
๐๐
๐=1
๐๐
๐ ๐, ๐
UNIVERSIDADE DE
SรO PAULO
๐๐: number of
distinct gray levelsin the quantized image
โCorrelation
๐3 =ฯ๐ฯ๐ ๐๐ ๐ ๐, ๐ โ ๐๐ฅ ๐๐ฆ
๐๐ฅ ๐๐ฆ
UNIVERSIDADE DE
SรO PAULO
14 texture featuresof Haralick et al.
14 texture featuresof Haralick et al.
โ ๐๐ฅ , ๐๐ฆ means
โ๐๐ฅ , ๐๐ฆ standard deviations
UNIVERSIDADE DE
SรO PAULO
๐๐ฆ ๐ =
๐=1
๐๐
๐ ๐, ๐๐๐ฅ ๐ =
๐=1
๐๐
๐ ๐, ๐
โ Sum of squares: Variance
๐4 =
๐
๐
๐ โ ๐ 2 ๐(๐, ๐)
UNIVERSIDADE DE
SรO PAULO
14 texture featuresof Haralick et al.
โ Inverse difference moment
๐5 =
๐
๐
1
1 + ๐ โ ๐ 2๐(๐, ๐)
UNIVERSIDADE DE
SรO PAULO
14 texture featuresof Haralick et al.
โ Sum average
โ Sum variance
๐6 =
๐=2
2๐๐
๐ ๐๐ฅ+๐ฆ (๐)
๐7 =
๐=2
2๐๐
๐ โ ๐82 ๐๐ฅ+๐ฆ(๐)
UNIVERSIDADE DE
SรO PAULO
14 texture featuresof Haralick et al.
UNIVERSIDADE DE
SรO PAULO
14 texture featuresof Haralick et al.
๐๐ฅ+๐ฆ ๐ =
๐=1
๐๐
๐=1
๐๐
๐(๐, ๐) ๐ = 2,3, โฆ , 2๐๐
๐ = ๐ + ๐
โ Sum entropy
โ Entropy
๐8 = โ
๐=2
2๐๐
๐๐ฅ+๐ฆ(๐) log ๐๐ฅ+๐ฆ ๐
๐9 = โ
๐
๐
๐ ๐, ๐ log ๐ ๐, ๐
UNIVERSIDADE DE
SรO PAULO
14 texture featuresof Haralick et al.
โDifference variance
โDifference entropy
๐10 = variance of ๐๐ฅโ๐ฆ
๐11 = โ
๐=0
๐๐โ1
๐๐ฅโ๐ฆ(๐) log ๐๐ฅโ๐ฆ(๐)
UNIVERSIDADE DE
SรO PAULO
14 texture featuresof Haralick et al.
UNIVERSIDADE DE
SรO PAULO
14 texture featuresof Haralick et al.
๐๐ฅโ๐ฆ ๐ =
๐=1
๐๐
๐=1
๐๐
๐(๐, ๐) ๐ = 0,1, โฆ , ๐๐ โ 1
๐ = ๐ โ ๐
โ Information measures of correlation 1
โ Information measures of correlation 2
๐12 =๐ป๐๐ โ ๐ป๐๐1
max ๐ป๐,๐ป๐
๐13 = 1 โ exp โ2 ๐ป๐๐2 โ ๐ป๐๐ 1/2
UNIVERSIDADE DE
SรO PAULO
14 texture featuresof Haralick et al.
๐ป๐๐ = โ
๐
๐
๐ ๐, ๐ log ๐ ๐, ๐
๐ป๐ and ๐ป๐ are entropy of ๐๐ฅ and ๐๐ฆ
๐ป๐๐1 = โ
๐
๐
๐ ๐, ๐ log ๐๐ฅ ๐ ๐๐ฆ ๐
๐ป๐๐2 = โ
๐
๐
๐๐ฅ ๐ ๐๐ฆ ๐ log ๐๐ฅ ๐ ๐๐ฆ ๐
UNIVERSIDADE DE
SรO PAULO
14 texture featuresof Haralick et al.
โMaximal correlation coefficient
๐14 = (second largest eigenvalue of ๐)1/2
where ๐ ๐, ๐ = ฯ๐๐ ๐,๐ ๐(๐,๐)
๐๐ฅ ๐ ๐๐ฆ(๐)
UNIVERSIDADE DE
SรO PAULO
14 texture featuresof Haralick et al.
Shape features
โCompactness ๐ถ๐
Perimeter PVertebral area A
UNIVERSIDADE DE
SรO PAULO
,4
12P
ACo
โ=
โ Fourier-descriptor-based feature FDF
Shape features
โ
=
โ=
1
0
2exp)(
1)(
N
n
nkN
jnzN
kZ
k = -N/2+1, โฆ, -1, 0, 1, 2, โฆ, N/2
z(n) = x(n) + j y(n)
n = 0, 1, ..., N-1
UNIVERSIDADE DE
SรO PAULO
โ Fourier-descriptor-based feature FDF
Shape features
+โ=
=
โ=
+โ+
=2
12
2
2
1
1
12
22
|)(|
|)(||)(|
N
Nk
N
kk
kk
N
kZ
kZkZFDF
UNIVERSIDADE DE
SรO PAULO
Shape features
โConvex deficiency CD
Vertebral area VA
๐ถ๐ท =๐ถ๐ป โ ๐๐ด
๐๐ด
UNIVERSIDADE DE
SรO PAULO
Convex hull CH
Shape features
โ 7 Central invariant moments (Hu)
๐1 = ยต20 + ยต02
๐2 = (ยต20 โ ยต02)2+4ยต11
2
๐3 = (ยต30 โ 3ยต12)2+(3ยต21 โ ยต03)
2
๐4 = (ยต30 + ยต12)2+(ยต21 + ยต03)
2
UNIVERSIDADE DE
SรO PAULO
Shape features
๐5
= ยต30 โ 3ยต12 ยต30 + ยต12 [ ยต30 + ยต122โ3 ยต21 + ยต03
2]+ (3ยต21 โ ยต03)(ยต21 + ยต03)[3(ยต30 + ยต12)
2โ(ยต21 + ยต03)2]
๐6
= ยต20 โ ยต02 (ยต30 + ยต12)2 โ (ยต21 + ยต03)
2
+ 4ยต11 ยต30 + ยต12 ยต21 + ยต03
๐7
= (3ยต21 โ ยต03)(ยต30 + ยต12)[(ยต30 + ยต12)2โ3(ยต21 + ยต03)
2] โ (ยต303ยต )( + )[3( + )2 ( + )2]
UNIVERSIDADE DE
SรO PAULO
Shape features
ยต00 = ๐00 = ยต
ยต10 = ยต01 = 0
ยต20 = ๐20 โ ยต๐ฅยฒ
ยต11 = ๐11 โ ยต๐ฅ๐ฆ
ยต02 = ๐02 โ ยต๐ฆยฒ
ยต30 = ๐30 โ 3๐20๐ฅ + 2ยต๐ฅยณ
ยต21 = ๐21 โ๐20๐ฆ โ 2๐11๐ฅ + 2ยต๐ฅยฒ๐ฆ
ยต12 = ๐12 โ๐02๐ฅ โ 2๐11๐ฆ + 2ยต๐ฅ ๐ฆยฒ
ยต03 = ๐03 โ 3๐02๐ฆ + 2ยต๐ฆยณ
UNIVERSIDADE DE
SรO PAULO
Shape features
๐๐๐
=
๐
๐
๐๐๐๐๐๐๐ ๐, ๐ , ๐, ๐ = 0,1,2, โฆ
๐ฅ =๐10
๐00๐ฆ =
๐01
๐00
UNIVERSIDADE DE
SรO PAULO
Organization of thefeature vector
Coefficientof variation
Skewness Kurtosis ... M7
1 2 3 ... 27
UNIVERSIDADE DE
SรO PAULO
Files of features
L1
L2
L3
L4
L5
txt files
UNIVERSIDADE DE
SรO PAULO
Inserting thereference classification
โManual addition of the class
โClassification according to radiologist andbiopsy
ClassCoefficientof variation
Skewness Kurtosis ... M7
1 2 3 ...
27
NormalVCF
Benign Malignant
UNIVERSIDADE DE
SรO PAULO
Feature selection
โ Software WEKA
โWrapper method for feature selection
โข kNN with k = 1, 3, ..., 13
โข Naรฏve Bayes
โข RBF network
โBest first as search method
โข Greedy search for the best subset of features
UNIVERSIDADE DE
SรO PAULO
Classification
โ Software WEKA
โClassifiers:
โข k-nearest neighbor: k = 1, 3, 5, 7, 9, 11, 13
โข Naรฏve Bayes
โข RBF network
โ Stratified 10-fold cross-validation
โข 9 folds for training, 1 fold for test
UNIVERSIDADE DE
SรO PAULO
Clinical Classes
โVCF vs Normal
โBenign VCF vs Malignant VCF
โMalignant VCF, Benign VCF, and Normal
UNIVERSIDADE DE
SรO PAULO
Validation
โConfusion Matrix
โข Sensitivity
โข Specificity
โข AUROC
โข% of correct classification
UNIVERSIDADE DE
SรO PAULO
๐ด๐ง and p-values UNIVERSIDADE DE
SรO PAULO
โ * for 0.01 โค p < 0.05
โ ** for 0.001 โค p < 0.01
โ *** for p < 0.001
โ p-values obtained using Wilcoxon rank-sum test
โ NS indicates no significant difference
โ NA indicates that ๐ด๐ง could not be obtained
๐ด๐ง and p-values UNIVERSIDADE DE
SรO PAULO
Benign VCF versus Malignant VCF
All VCFs together versus Normal
Feature Significance ๐จ๐ Significance ๐จ๐๐ถ๐ NS 0.580 *** 0.751
๐๐๐๐ค *** 0.861 * 0.549
๐พ๐ข๐๐ก *** 0.824 NS 0.532
๐ป1 *** 0.849 NS 0.625
๐ป2 *** 0.866 * 0.661
๐ป3 NS 0.480 NS 0.629
๐ป4 *** 0.874 NS 0.642
๐ป5 *** 0.844 * 0.577
๐ป6 *** 0.829 *** 0.731
๐ป7 *** 0.871 NS 0.640
๐ป8 *** 0.854 ** 0.620
๐ป9 *** 0.858 *** 0.647
๐ป10 *** 0.871 ** 0.674
๐ด๐ง and p-values UNIVERSIDADE DE
SรO PAULO
Benign VCF versus Malignant VCF
All VCFs together versus Normal
Feature Significance ๐จ๐ Significance ๐จ๐
H11 *** 0.868 ** 0.632
H12 *** 0.731 NS 0.524
H13 *** 0.854 *** 0.614
H14 NS 0.566 NS 0.462
Co *** 0.722 *** 0.864
FDF *** 0.837 NS 0.449
CD *** 0.700 *** 0.881
M1 NS 0.567 *** 0.964
M2 NS 0.518 *** 0.932
M3 ** 0.655 * 0.887
M4 * 0.617 NS 0.936
M5 NS 0.389 NS NA
M6 NS 0.480 NS 0.498
M7 NS 0.538 NS NA
๐ด๐ง and p-values UNIVERSIDADE DE
SรO PAULO
Mean and standard deviation of features UNIVERSIDADE DE
SรO PAULO
Mean and standard deviation of features UNIVERSIDADE DE
SรO PAULO
โMean skewness of malignant VCFs is higherthan that for benign VCFs
โข T1 signals are distributed more on the lowerside of the histogram for malignant VCFs
โ๐ป6 and ๐ป7 show large differences in their mean values for malignant VCFs versus benign VCFs
Mean and standard deviation of features UNIVERSIDADE DE
SรO PAULO
Feature selection UNIVERSIDADE DE
SรO PAULO
Feature selection UNIVERSIDADE DE
SรO PAULO
โ k-NN did not select the gray-level featuresfor benign vs malignant VCFs
โข ๐น๐ท๐น, ๐5, ๐ป10, and ๐ป13were selected at least three times
โCV is statistically significant for all VCFs vsnormal vertebral bodies and was selectedfor all classifiers
Feature selection UNIVERSIDADE DE
SรO PAULO
โVarious texture features were selected for both types of classification
โNaรฏve Bayes selected the highest number offeatures for both types of classification
Classification UNIVERSIDADE DE
SรO PAULO
Classifier ACC rate % AUROC
k-NN
k = 7 82.4 0.84
k = 9 81.4 0.90
k = 11 84.3 0.90
k = 13 84.3 0.90
Naรฏve Bayes
RBF network
85.3 0.92
78.4 0.86
Classifier ACC rate % AUROC
k-NN
k = 7 90.1 0.95
k = 9 89.0 0.92
k = 11 89.0 0.92
k = 13 89.5 0.94
Naรฏve Bayes
RBF network
90.6 0.97
91.1 0.94
โ Benign vs malignantVCFs
โ All VCFs vs normal vertebral bodies
Classification UNIVERSIDADE DE
SรO PAULO
โRBF network classifier for benign vsmalignant VCFs
โข ACC rate was the lowest obtained
โข AUROC is only better than that of 7-NN
โRBF network classifier for all VCFs vs normal vertebral bodies
โข ACC rate is the highest obtained
Classification UNIVERSIDADE DE
SรO PAULO
โAUROC for classification of all VCFs together vs normal vertebral bodies is at least 0.92
โAUROC of the naรฏve Bayes classifier is 0.97 for this purpose
โข Better than the previous study using only shapefeatures in which AUROC was 0.945
โ This shows the importance of texture andgray-level features for this purpose
Classification UNIVERSIDADE DE
SรO PAULO
โAUROC for classification of benign vs malignant VCFs is 0.92 for naรฏve Bayes
โข Better than the previous study in which thehighest AUROC was 0.91 for 3-NN
โ In a previous study using only shapefeatures the highest AUROC was 0.78
โข This shows the importance of texture and gray-level features for this purpose
Benign VCFs, malignant VCFs, and normal vertebral bodies UNIVERSIDADE DE
SรO PAULO
Predicted classification True classification
Malignant VCFs Benign VCFs Normal vertebral bodies
39 5 5 Malignant VCFs
13 35 5 Benign VCFs
4 1 84 Normal vertebral bodies
โข Features selected: โข CV, Skew, H2, H3, H5, H6, H8, H9, H11, H12,
H13,H14,Co, FDF, CD, M1, M3, and M7
โข Weighted average AUROC of 0.94
โข ACC rate of 82.7%
โManual segmentation of the vertebral bodies
โข Automatic segmentation methods could lead to the realization of a clinically useful CAD system
โ Individual and separate analysis of thevertebral bodies ignores importantinformation outside their regions
Limitations of the study UNIVERSIDADE DE
SรO PAULO
โ The use of only the median sagittal slice
โ Some lateral VCFs may be misclassified
โ Extension of segmentation and feature extraction methods to 3D is desirable
Limitations of the study UNIVERSIDADE DE
SรO PAULO
โAnalysis of only T1-weighted MRI
โข Benign VCFs
โข isointense vertebra in T2-weighted and T1-weighted MRI after gadolinium contrast
โขMalignant VCFs
โข heterogeneous or high signal in T2-weighted and in
T1-weighted MRI after gadolinium contrast
Limitations of the study UNIVERSIDADE DE
SรO PAULO
โMost of the features presented are important for both types of VCF classification
โ For benign vs malignant VCFs
โข AZ values of texture and gray-level features are higher than those shape features
โ For all VCFs vs normal vertebral bodies
โข AZ values of shape features are higher than those of texture and gray-level features
Conclusion UNIVERSIDADE DE
SรO PAULO
โ The features FDF and CV follow the
opposite trend
โ The naรฏve Bayes method was the best classifier in both types of classification
โ The proposed methods are promising
for CAD of VCFs
Conclusion UNIVERSIDADE DE
SรO PAULO
โ Future works:
โข Evaluate our methods with the inclusion of anautomatic segmentation method
โข Extend the methods to 3D analysis of vertebral bodies
Conclusion UNIVERSIDADE DE
SรO PAULO
โ Sรฃo Paulo Research Foundation (FAPESP)
โข 2014/12135-0 and 2015/08778-6
โ National Council of Technological and ScientificDevelopment (CNPq)
โ Natural Sciences and Engineering Research Council of Canada
โ Ph.D students
โข Rafael de Menezes-Reis
โข Faraz Oloumi
AcknowledgmentUNIVERSIDADE DE
SรO PAULO
Feature selection: benign vs malignant
VCFs
UNIVERSIDADE DE
SรO PAULO
Featurek-NN Naรฏve
Bayes
RBF
Networkk = 7 k = 9 k = 11 k = 13
X X
X X X
X X X X
X X X
X X X
X
X
X X X X X
X X
X X
X X X X X X
X
X
Feature selection: all VCFs
vs normal vertebral bodies
UNIVERSIDADE DE
SรO PAULO
Featurek-NN Naรฏve
Bayes
RBF
Networkk = 7 k = 9 k = 11 k = 13
X X X X X X
X
X
X
X X
X
X
X
X
X
X X X X
X
X X X X X X
X X X
X X X X
X X X X