Classification of Benign and Malignant Vertebral Compression Fractures in Magnetic Resonance Images

Transcription

1 Classification of Benign and Malignant Vertebral Compression Fractures in Magnetic Resonance Images Lucas Frighetto-Pereira, Guilherme Augusto Metzner, Paulo Mazzoncini de Azevedo-Marques, Rangaraj Mandayam Rangayyan, Marcello Henrique Nogueira-Barbosa

2 Anatomy of the spine MRI sagittal slice Vertebra Intervertebral Disc Vertebral Arch Lumbar Spine Vertebral Body

3 Vertebral compression fractures (VCFs) Partial collapse of vertebral bodies Traumatic VCFs raise no doubt about their etiology But a recent vertebral collapse without history of significant trauma creates difficulty in defining the cause of the VCF

4 Medical diagnosis Young patient with a VCF History of significant acute trauma Usually easy diagnosis

5 Medical diagnosis Elderly patient with VCF No history of significant acute trauma Diagnosis?

6 VCFs without history of significant trauma VCFs are the most common type of osteoporotic fractures The elderly have a high incidence of VCFs related to metastatic cancer affecting bone MRI is the most commonly used imaging method for spinal diseases and early detection of fractures

7 T1-Weighted MRI Osteoporotic VCF Metastatic VCF

8 Clinical classification of VCFs Osteoporotic VCFs classified as Benign VCFs Metastatic VCFs classified as Malignant VCFs

9 Benign VCFs in T1-weighted MRI Partial preservation of normal fatty bonemarrow signal in the vertebral body Degeneration of normally rectangular shapes of vertebrae into concave and rough shapes with indentations Rougher contours than malignant VCFs and normal vertebrae

10 Malignant VCFs in T1-weighted MRI Global reduction of signal intensity or nodular abnormality in the affected vertebral body Could result in a posterior convexity without substantial concavities May also cause the contours of vertebrae to be relatively smoothened due to convexity

11 Normal Benign VCFs Malignant VCFs

12 Benign vs Malignant VCFs Both tend to create concavities in the vertebral plateaus Could cause doubt in the diagnosis Correct classification is critical for planning treatment

13 Benign VCF Malignant VCF Which image has the malignant VCF and which one has the benign VCF?

14 Objectives Study the characteristics of VCFs in MRI Develop image processing techniques to extract features Classify VCFs Normal Benign Fractured Malignant

15 Study steps Selection of cases and images Manual segmentation of vertebral bodies Extraction of features of vertebral bodies Classification, validation, and statistical analysis

16 Database University Hospital of Ribeirão Preto Medical School University of São Paulo Cases and images collected from the Radiology Information System (RIS) Cases from September 2010 to March 2014 Philips 1.5T MRI System T1-weighted MRI

17 Database Lumbar vertebral bodies (L1 to L5) Median sagittal slice TIFF images with 8-bits/pixel 153 exams analyzed, 63 selected 38 women, 25 men Mean age: 62 years

18 Database 63 selected exams: At least one VCF per patient The nonfractured vertebral bodies of patients without malignant fractures are considered to be normal

19 Excluded cases Vertebral fractures secondary to trauma Infection and avascular necrosis Severe degenerative scoliosis Previous surgeries, radiotherapy, and chemotherapy

20 Database L5 L4 L3 L2 L1 Total Benign VCFs Malignant VCFs Normal Total

21 Examples of vertebral bodies Normal Benign VCFs Malignant VCFs

22 Manual segmentation MRI exam Vertebral body masks

23 Software flow chart

24 MRI exam and its mask

25 Detection of the coordinates of the vertebral bodies L1 L3 L2 L4 L5

26 Normalization of the MR images DiscsMean Extraction of blocks of intervertebral discs using the mask ROIs as reference. 5x5 disc block

27 Normalization of the MR images newimg i, j = imgoriginal(i, j) discsmean imgnorm i, j = 255 newimg i, j min newimg max newimg min newimg

28 MRI exam Mask Processing new image...

29 Detection of the ROIs

30 ROIs of the vertebral bodies Normal Benign VCFs Malignant VCFs

31 Computation of the features 3 Statistical gray-level features 14 Texture features 10 Shape features 27 Features

32 Statistical gray-level features Coefficient of variation Skewness Kurtosis

33 Statistical gray-level features Coefficient of variation (CV ) μ = i=1 x i CV = σ μ σ = i=1 x i μ 2

34 Statistical gray-level features Skewness skewness = (x i μ) σ 3 i=1

35 Statistical gray-level features Kurtosis kurtosis = (x i μ) σ 4 i=1

36 Differences in texture between normal and VCFs Malignant VCF Normal Malignant VCF Benign VCF Malignant VCF Malignant VCF

37 Texture features Gray-level cooccurrence matrix 14 texture features of Haralick et al.

38 Cooccurrence matrix Ex: Image 5x5 pixels, 3 gray levels Cooccurrence matrix p i, j for i = 0, j = Distance = 1 pixel Angle = ± Number of pixels of intensity 0 that are at ±45 degrees and distance 1 of pixels of intensity 2

39 14 texture features of Haralick et al. Angular second moment (Energy) f 1 = p(i, j) 2 Contrast i j Ng 1 Ng Ng f 2 = n 2 p i, j n=0 i=1 j=1 Ng: number of distinct gray levels in the quantized image

40 14 texture features of Haralick et al. Correlation f 3 = σ i σ j ij p i, j μ x μ y σ x σ y

41 14 texture features of Haralick et al. μ x, μ y means σ x, σ y standard deviations Ng p x i = p i, j j=1 Ng p y j = i=1 p i, j

42 14 texture features of Haralick et al. Sum of squares: Variance f 4 = i j i μ 2 p(i, j)

43 14 texture features of Haralick et al. Inverse difference moment f 5 = i j i j 2 p(i, j)

44 14 texture features of Haralick et al. Sum average 2Ng f 6 = i=2 i p x+y (i) Sum variance 2Ng f 7 = i=2 i f 8 2 p x+y (i)

45 14 texture features of Haralick et al. Ng Ng p x+y k = i=1 j=1 k = i + j p(i, j) k = 2,3,, 2Ng

46 14 texture features of Haralick et al. Sum entropy 2Ng f 8 = i=2 p x+y (i) log p x+y i Entropy f 9 = p i, j log p i, j i j

47 14 texture features of Haralick et al. Difference variance f 10 = variance of p x y Difference entropy Ng 1 f 11 = i=0 p x y (i) log p x y (i)

48 14 texture features of Haralick et al. Ng Ng p x y k = p(i, j) k = 0,1,, Ng 1 i=1 j=1 k = i j

49 14 texture features of Haralick et al. Information measures of correlation 1 f 12 = HXY HXY1 max HX, HY Information measures of correlation 2 f 13 = 1 exp 2 HXY2 HXY 1/2

50 14 texture features of Haralick et al. HXY = p i, j log i j p i, j HX and HY are entropy of p x and p y HXY1 = p i, j log i j p x i p y j HXY2 = p x i p y j log p x i p y j i j

51 14 texture features of Haralick et al. Maximal correlation coefficient f 14 = (second largest eigenvalue of Q) 1/2 where Q i, j = σ k p i,k p(j,k) p x i p y (k)

52 Shape features Compactness C o 4 A C o = 1 2 P, Perimeter P Vertebral area A

53 Shape features Fourier-descriptor-based feature FDF Z ( k) 1 N 1 = N n= 0 z( n)exp j 2 N nk k = -N/2+1,, -1, 0, 1, 2,, N/2 z(n) = x(n) + j y(n) n = 0, 1,..., N-1

54 Fourier-descriptor-based feature FDF Shape features + = = = + + = ) ( ) ( ) ( N N k N k k k k N k Z k Z k Z FDF

55 Shape features Convex deficiency CD CD = CH VA VA Convex hull CH Vertebral area VA

56 Shape features 7 Central invariant moments (Hu) M 1 = µ 20 + µ 02 M 2 = (µ 20 µ 02 ) µ 11 M 3 = (µ 30 3µ 12 ) 2 +(3µ 21 µ 03 ) 2 M 4 = (µ 30 + µ 12 ) 2 +(µ 21 + µ 03 ) 2

57 Shape features M 5 = µ 30 3µ 12 µ 30 + µ 12 [ µ 30 + µ µ 21 + µ 03 2 ] + (3µ 21 µ 03 )(µ 21 + µ 03 )[3(µ 30 + µ 12 ) 2 (µ 21 + µ 03 ) 2 ] M 6 = µ 20 µ 02 (µ 30 + µ 12 ) 2 (µ 21 + µ 03 ) 2 + 4µ 11 µ 30 + µ 12 µ 21 + µ 03 M 7 = (3µ 21 µ 03 )(µ 30 + µ 12 )[(µ 30 + µ 12 ) 2 3(µ 21 + µ 03 ) 2 ] (µ 30

58 Shape features µ 00 = m 00 = µ µ 10 = µ 01 = 0 µ 20 = m 20 µx² µ 11 = m 11 µxy µ 30 = m 30 3m 20 x + 2µx³ µ 21 = m 21 m 20 y 2m 11 x + 2µx²y µ 12 = m 12 m 02 x 2m 11 y + 2µx y² µ 03 = m 03 3m 02 y + 2µy³ µ 02 = m 02 µy²

59 Shape features m pq = i j i p j q img i, j, p, q = 0,1,2, x = m 10 m 00 y = m 01 m 00

60 Organization of the feature vector Coefficient of variation Skewness Kurtosis... M7

61 Files of features L1 L2 L3 txt files L4 L5

62 Inserting the reference classification Manual addition of the class Classification according to radiologist and biopsy Coefficient of variation Skewness Kurtosis... M7 Class VCF Normal Benign Malignant

63 Feature selection Software WEKA Wrapper method for feature selection knn with k = 1, 3,..., 13 Naïve Bayes RBF network Best first as search method Greedy search for the best subset of features

64 Classification Software WEKA Classifiers: k-nearest neighbor: k = 1, 3, 5, 7, 9, 11, 13 Naïve Bayes RBF network Stratified 10-fold cross-validation 9 folds for training, 1 fold for test

65 Clinical Classes VCF vs Normal Benign VCF vs Malignant VCF Malignant VCF, Benign VCF, and Normal

66 Validation Confusion Matrix Sensitivity Specificity AUROC % of correct classification

67 A z and p-values * for 0.01 p < 0.05 ** for p < 0.01 *** for p < p-values obtained using Wilcoxon rank-sum test NS indicates no significant difference NA indicates that A z could not be obtained

68 A z and p-values Benign VCF versus Malignant VCF All VCFs together versus Normal Feature Significance A z Significance A z CV NS *** Skew *** * Kurt *** NS H 1 *** NS H 2 *** * H 3 NS NS H 4 *** NS H 5 *** * H 6 *** *** H 7 *** NS H 8 *** ** H 9 *** *** H 10 *** ** 0.674

69 A z and p-values Benign VCF versus Malignant VCF All VCFs together versus Normal Feature Significance A z Significance A z H 11 *** ** H 12 *** NS H 13 *** *** H 14 NS NS C o *** *** FDF *** NS CD *** *** M 1 NS *** M 2 NS *** M 3 ** * M 4 * NS M 5 NS NS NA M 6 NS NS M 7 NS NS NA

70 A z and p-values

71 Mean and standard deviation of features

72 Mean and standard deviation of features Mean skewness of malignant VCFs is higher than that for benign VCFs T1 signals are distributed more on the lower side of the histogram for malignant VCFs H 6 and H 7 show large differences in their mean values for malignant VCFs versus benign VCFs

73 Mean and standard deviation of features

74 Feature selection

75 Feature selection k-nn did not select the gray-level features for benign vs malignant VCFs FDF, M 5, H 10, and H 13 were selected at least three times CV is statistically significant for all VCFs vs normal vertebral bodies and was selected for all classifiers

76 Feature selection Various texture features were selected for both types of classification Naïve Bayes selected the highest number of features for both types of classification

77 Classification Benign vs malignant VCFs Classifier ACC rate % AUROC k = k-nn k = k = k = Naïve Bayes RBF network All VCFs vs normal vertebral bodies Classifier ACC rate % AUROC k = k-nn k = k = Naïve Bayes RBF network k =

78 Classification RBF network classifier for benign vs malignant VCFs ACC rate was the lowest obtained AUROC is only better than that of 7-NN RBF network classifier for all VCFs vs normal vertebral bodies ACC rate is the highest obtained

79 Classification AUROC for classification of all VCFs together vs normal vertebral bodies is at least 0.92 AUROC of the naïve Bayes classifier is 0.97 for this purpose Better than the previous study using only shape features in which AUROC was This shows the importance of texture and gray-level features for this purpose

80 Classification AUROC for classification of benign vs malignant VCFs is 0.92 for naïve Bayes Better than the previous study in which the highest AUROC was 0.91 for 3-NN In a previous study using only shape features the highest AUROC was 0.78 This shows the importance of texture and graylevel features for this purpose

81 Benign VCFs, malignant VCFs, and normal vertebral bodies Predicted classification Malignant VCFs Benign VCFs Normal vertebral bodies True classification Malignant VCFs Benign VCFs Normal vertebral bodies Features selected: CV, Skew, H2, H3, H5, H6, H8, H9, H11, H12, H13,H14,Co, FDF, CD, M1, M3, and M7 Weighted average AUROC of 0.94 ACC rate of 82.7%

82 Limitations of the study Manual segmentation of the vertebral bodies Automatic segmentation methods could lead to the realization of a clinically useful CAD system Individual and separate analysis of the vertebral bodies ignores important information outside their regions

83 Limitations of the study The use of only the median sagittal slice Some lateral VCFs may be misclassified Extension of segmentation and feature extraction methods to 3D is desirable

84 Limitations of the study Analysis of only T1-weighted MRI Benign VCFs isointense vertebra in T2-weighted and T1-weighted MRI after gadolinium contrast Malignant VCFs heterogeneous or high signal in T2-weighted and in T1-weighted MRI after gadolinium contrast

85 Conclusion Most of the features presented are important for both types of VCF classification For benign vs malignant VCFs AZ values of texture and gray-level features are higher than those shape features For all VCFs vs normal vertebral bodies AZ values of shape features are higher than those of texture and gray-level features

86 Conclusion The features FDF and CV follow the opposite trend The naïve Bayes method was the best classifier in both types of classification The proposed methods are promising for CAD of VCFs

87 Conclusion Future works: Evaluate our methods with the inclusion of an automatic segmentation method Extend the methods to 3D analysis of vertebral bodies

88 Acknowledgment São Paulo Research Foundation (FAPESP) 2014/ and 2015/ National Council of Technological and Scientific Development (CNPq) Natural Sciences and Engineering Research Council of Canada Ph.D students Rafael de Menezes-Reis Faraz Oloumi

89 Feature k-nn k = 7 k = 9 k = 11 k = 13 Naïve Bayes RBF Network X X X X X X X X X Feature selection: benign vs malignant VCFs 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

90 Feature k-nn k = 7 k = 9 k = 11 k = 13 Naïve Bayes RBF Network X X X X X X X X X X X X X Feature selection: all VCFs vs normal vertebral bodies X X X X X X X X X X X X X X X X X X X X X