4. •Lung cancer is the leading cause of cancer deaths.
•Most patients diagnosed with lung cancer already have advanced disease
–40% are stage IV and 30% are III
–The current five-year survival rate is only 16%
•Defective nodules are detected at an early stage
–The survival rate can be increased
4
5. 5
(a) male (b) female
Trends in death rates for selected cancers, United States, 1930-2008
6. •Early detection of lung nodules is extremely important for the diagnosis and clinical management of lung cancer
•Lung cancer had been commonly detected and diagnosed on chest radiography
•Since the early 1990s CT has been reported to improve detection and characterization of pulmonary nodules
6
7. •CT was introduced in 1971
–Sir Godfrey Hounsfield, United Kingdom
•CT utilize computer-processed X-rays
–to produce tomographic images or 'slices' of specific areas of the body
•The Hounsfield unit (HU) scale is a linear transformation of the original linear attenuation coefficient measurement into one in which the radio density of distilled water
7
waterwaterx1000 HU
8. 8
The HU of common substances
Substance
HU
Air
−1000
Lung
−500
Fat
−84
Water
0
Cerebrospinal Fluid
15
Blood
+30 to +45
Muscle
+40
Soft Tissue, Contrast Agent
+100 to +300
Bone
+700(cancellous bone)to +3000 (dense bone)
Nodule
9. •Lung cancer screening is currently implemented using low-dose CT examinations
•Advanced in CT technology
–Rapid image acquisition with thinner image sections
–Reduced motion artifacts and improved spatial resolution
•The typical examination generates large-volume data sets
•These large data sets must be evaluated by a radiologist
–A fatiguing process
9
10. •The use of pulmonary nodule detection CAD system can provide an effective solution
•CAD system can assist radiologists by increasing efficiency and potentially improving nodule detection
10
General structure of pulmonary nodule detection system
11. CAD systems
Lung segmentation
Nodule Candidate Detection
False Positive Reduction
Suzuki et al.(2003)[26]
Thresholding
Multiple thresholding
MTANN
Rubin et al.(2005)[27]
Thresholding
Surface normal overlap
Lantern transform and rule-based classifier
Dehmeshki et al.(2007)[28]
Adaptive thresholding
Shape-based GATM
Rule-based filtering
Suarez-Cuenca et al.(2009)[29]
Thresholding and 3-D connected component labeling
3-D iris filtering
Multiple rule-based LDA classifier
Golosio et al.(2009)[30]
Isosurface-triangulation
Multiple thresholding
Neural network
Ye et al.(2009)[31]
3-D adaptive fuzzy segmentation
Shape based detection
Rule-based filtering and weighted SVM classifier
Sousa et al.(2010)[32]
Region growing
Structure extraction
SVM classifier
Messay et al.(2010)[33]
Thresholding and 3-D connected component labeling
Multiple thresholding and morphological opening
Fisher linear discriminant and quadratic classifier
Riccardi et al.(2011)[34]
Iterative thresholding
3-D fast radial filtering and scale space analysis
Zernike MIP classification based on SVM
Cascio et al.(2012)[35]
Region growing
Mass-spring model
Double-threshold cut and neural network
11
12. •To evaluate the performance of the proposed method, Lung Image Database Consortium (LIDC) database is applied
•LIDC database, National Cancer Institute (NCI), United States
–The LIDC is developing a publicly available database of thoracic computed tomography (CT) scans as a medical imaging research resource to promote the development of computer-aided detection or characterization of pulmonary nodules
•The database consists of 84 CT scans
–100-400 Digital Imaging and Communication (DICOM) images
–An XML data file containing the physician annotations of nodules
–148 nodules
–The pixel size in the database ranged from 0.5 to 0.76 mm
–The reconstruction interval ranged from 1 to 3mm
12
15. 15
•Air has an attenuation of -1000 HU
•Most lung tissue is in the range of -910 HU to -500 HU
•The chest wall, blood vessel, and bone are above -500 HU
•The low and high intensities are differentiable around the intensity -500 HU
(,,)(,,)500ixyzIxyzHUS
16. 16
Input CT images, their intensity histograms, and thresholded images
17. •A fixed threshold is applicable to segment lung area
–The intensity ranges of images are varied by different acquisition protocols
•To obtain optimal threshold
–Iterative approach continues until the threshold converges
–The initial threshold :
– is i th threshold and new threshold as
17
(0)500THU (1) 2iobT ()iT
18. 18
Input CT images, their intensity histograms, and thresholded images
19. •White areas
–non-body voxels
–including lung cavity
•Black areas
–body voxels
–excluding lung region
•Lung regions are extracted from the non- body voxels by using 3- D connected component labeling
19
18-connectivity voxels
21. •To extract lung volume
–Remove rim attached to boundaries of image
–The first and the second largest volumes are selected as the lung region
•The lung region contains small holes
–To remove these holes
–Morphological hole filling operations are applied
21
|lungfirstsecondSll
22. 22
Binary images of the selected lung region
Lung mask images after hole filling
23. •The contour of the lung volume is needed to correct
–To include wall side nodule (juxta-pleural nodule)
23
Extracted lung region using 3D connected component labeling and contour corrected lung region (containing wall side nodule)
30. •Detection of nodule candidates is important
•The performance of nodule detection system relies on the accuracy of candidate detection
•ROI extraction
–Optimal multi-thresholding
•Nodule candidates detection and segmentation
–Rule-based pruning
30
31. •The traditional multi-thresholding method needs many steps of grey levels
•An iterative approach is applied to select the threshold value
•The optimal threshold value is calculated on median slice of lung CT scan
31
(1) 2iobT
32. •The optimal threshold value
–A base threshold for multi-thresholding
•Additional six threshold values are obtained
–Base threshold + 400,+ 300,+ 200,+ 100, - 100, and - 200
32
33. •Rule based classifier removes vessels and noise
•Vessel removing
–Volume is extremely bigger than nodule
–Elongated object
•Noise removing
–Radius of ROI is smaller than 3mm
–Bigger than 30mm
•Remaining ROIs are nodule candidates
33
34. 34
Rule
Description
R1
Small noise
R2
Vessel
R3
Large noise
R4
Nodule
Pruning rules for nodule candidate detection
36. 36
(d) (e) (f) The results of nodule candidate detection: (a,d) ROIs, (b,e) vessel, and (c,f) nodule candidates after rule-based pruning
(a) (b) (c)
37. •The features are useful information that describe characteristics of the nodule candidates
•In the proposed CAD system, these features will be used to train the GPC
•The proposed feature extraction process consists of two stages
–The variety types of features are extracted from the nodule candidates
–Subsets of features are selected and combined into sub-groups
37
38. 38
Index Feature Index Feature
2-D geometric features Mean inside
Area Mean outside
Diameter Variance inside
Perimeter Skewness inside
Circularity Kurtosis inside
3-D geometric features Eigenvalues
Volume 3-D intensity based statistical features
Compactness Minimum value inside
Bounding Box Dimensions Mean inside
Principal Axis Length Mean outside
Elongation Variance inside
2-D intensity based statistical features Skewness inside
Minimum value inside Kurtosis inside
1 f
2 f
3 f
4 f
5 f
6 f
7 9 f ~ f
10 12 f ~ f
13 f
14 f
15 f
16 f
17 f
18 f
19 f
20 27 f ~ f
28 f
29 f
30 f
31 f
32 f
33 f
Features for nodule detection
39. Feature vector Description
2-D geometric features
3-D geometric features
2-D intensity-based statistical features
3-D intensity-based statistical features
2-D features
3-D features
Geometric features
Intensity-based statistical features
All features
39
1 1 4 f { f ,..., f }
2 5 13 f { f ,..., f }
3 14 27 f { f ,..., f }
4 28 33 f { f ,..., f }
5 1 3 f f f
6 2 4 f f f
7 1 2 f f f
8 3 4 f f f
1 2 3 4 f f f f f
Eight different groups of feature vectors
40. •Genetic Programming (GP)
–An evolutionary optimization technique
•The basic structure of GP is very similar to Genetic Algorithm(GA)
•The chromosome
–GA : variable (binary digit or string)
–GP : program (tree or graph)
40
A function represented as a tree structure
41. •GP chromosome
–The terminal set
•The elements of feature vector extracted from nodule candidate images
•Randomly generated constants with in the range 0,1
–The function set
•Four standard arithmetic operator namely plus, minus, multiply and division
•Additional mathematical operators log, exp, abs, sin and cos
•All operators in the function set are protected to avoid exception
•GP evolves combination of the terminal set and function set
41
42. • Fitness Function
– evaluate every individuals in GP generation
• True positive rate (TPR)
• Specificity (SPC)
– SPC is the value subtracted from 1 to FPR and also called true negative
rate (TNR)
• Area under the ROC curve (AUC)
– ROC curve is plotted between TP and FP for different threshold values
– AUC is area under the ROC curve and a good measure of classifier
performance in different condition
42
TP
TPR
TP FN
1 1
TN FP
SPC FPR
TN FP FP TN
f TPRFPR AUC
43. 43
Objective
To evolve a optimum classifier with a maximum TPR, SPC and AUC
Function Set
+,-,*,protected division, log, exp, abs, sin and cos
Terminal Set
Elements of a feature vector and randomly generated constants
Fitness
Fit(B)=TPR×SPC×AUC
Selection
Generational
Wrapper
Positive if , else negative
Population Size
300
Generation Size
80
Initial Tree Depth Limit
6
Initial population
Ramped half and half
GP Operators prob.
Variable ratio of crossover mutation is used
Sampling
Tournament
Survival mechanism
Keep the best individuals
Real max. tree level
30
54. •Coarse to fine hierarchical block-image analysis
–Block size : 32, 24, 16, 12, 8
•3-D CT scan is split into 3- D block-images
•The non-informative block-images are filtered out by using entropy analysis
54
55. 55
Result images after block splitting with respect to various block sizes
56. • Calculate the entropy H(x) on block image
• Select informative blocks by using entropy
56
1
2 2
1
1
( ) ( ) log ( ) log ( )
( )
n n
i i
H x p i p i p i
p i
57. 57
The entropy histograms of block-images for five different block sizes
(x-axis : entropy value, y-axis : number of blocks, (a) 32, (b) 24, (c) 16, (d) 12, and (e) 8)
58. •The selected block- image is enhanced
•The object in the selected block-image is segmented
•The location of block image is adjusted
58
59. 59
•Block-image enhancement is presented for more accurate analysis
•3-D coherence-enhancing diffusion (CED) filter
–Hessian matrix based
–Preserve small spherical structure (nodule)
–Enhance tubular structure (vessel)
(a) Input image and (b) the result image after enhancement
61. •The location of block-image should be adjusted
–The segmented object is not located in the center of the block
•Block location is iteratively updated by using centroid of the segmented object
•The iteration of the adjustment continues until the center position converges
•Or distance between the adjusted location and the original location is larger than half of the block size
61
62. 62
Iterations of automatic block location adjustment, upper: 3-D shapes, lower: the median slices of 3-D block; (a) the first; (b) the fifth; and (c) the last iterations of adjustment
63. 63
•Three different types of features are extracted from nodule candidate block- images
•Nodule has their own shapes
–Important characteristics to distinguish
•2-D and 3-D geometric features describe the shape of nodule candidates
65. •Support vector machine (SVM)
–SVM is a useful technique for data classification
–Supervised learning models with associated learning algorithms
–SVM analyze data and recognize patterns
–Classification and regression analysis
65
66. •The basic SVM takes a set of input data and predicts two possible classes for each given input
•Training dataset
•The SVM requires the solution of the following optimization problem
66
67. •SVM can efficiently perform non-linear classification using the kernel trick
•Kernel function
–Polynomial function
–Radial basis function
–Minkowski distance function
67
68. •k-fold cross-validation is applied to evaluated the proposed classifier
•Performance validation measure
–The number of true positives (TPs) and false positives (FPs)
–Accuracy, sensitivity, specificity, and area under the ROC curve.
68
69. 69
k
p
AUC
Accuracy
Sensitivity
Specificity
5
0.25
0.9738
91.52%
87.16%
95.88%
7
0.25
0.9784
93.97%
91.02%
96.92%
10
0.25
0.9736
92.43%
88.97%
95.88%
The k-fold cross validation results of SVM classifiers with radial basis function kernel for different k values
71. 71
ROC curves of the SVM classifiers with respect to three different kernel functions, SVM-r: radial basis function, SVM-p: polynomial function, and SVM-m: Minkowski distance function; (a) p = 0:25 and (b) p = 1.
74. •Eigenvalue decomposition of Hessian Matrix
–Dot enhancement filter
–Feature extraction
•Multi-scale dot enhancement filter
–Enhance the nodules
–The shape of nodules is like dot or ball
74
75. 75
•Multi-scale dot enhancement filter is based on eigenvalue of Hessian matrix
•Hessian Matrix
•Local structure information is obtained by Hessian matrix
76. •Eigenvalue decomposition of Hessian Matrix
–Structure information : surface-ness, curve-ness, and point- ness
–This information is expressed in the three singular tensors (stick, plate, and ball)
•Tensor based representation
76
79. •The dot enhancement filter is applied to enhance the spherical object, such as nodule
•For each voxel, the dot value is defined as
• are three eigenvalues from the Hessian matrix
•Gaussian image smoothing with a variety scales is performed prior to the calculation of the gradient for different size of nodules and reducing noise
79
80. •Assuming that the diameter of nodule to be detected are in a range the N discrete smoothing scales ___ in the range of can be calculated as
where and each scale has corresponding nodule diameter
•The maximum dot value calculated among the different smoothing scales
•Five steps smoothing scales are used in the range of nodule diameter [3mm, 30mm]
80 (1/(1)) 10(/)Nrdd
81. •The image block is extracted as a potential nodule candidate
–The dot values are larger than predefined threshold
•The dimension of the image block is
•It is noted that the size of the image block is considered at the relation to the corresponding smoothing scale as follows:
where the braces indicate the ceiling function
81
82. •A novel shape-based feature extraction method is proposed
•Angular Histogram of Surface Normal Feature
•The feature extraction has important role in the pulmonary nodule CAD system
•The detected nodule candidates are considered as nodules or non-nodules using the extracted feature information
82
83. •Popular approach in the last decade for 2-D images
•The scale invariant feature transform (SIFT)
–It can extract salient points and feature descriptors in the most invariant way with respect to scaling, translation, orientation, affine changes and illumination within images
–The SIFT is designed and tested on 2-D images of 3-D object.
–Allaire et al. proposed fully orientation invariant 3-D SIFT
•The histograms of oriented gradients (HOG)
–Describing salient points on 2-D images of 3-D objects
–Scherer et al. proposed the 3-D extension of HOG is proposed for 3-D object retrieval
83
84. •The shape-based feature descriptor is extracted for small 3-D object in image patch
•The AHSN feature extraction method is proposed to analyze the shape of the target object
•The eigenvalue decomposition of the Hessian matrix is applied to every voxels for target image
•The histograms are obtained on surface-ness information
–surface saliency :
–surface normal vector :
84
85. •The orientation of surface normal vector is obtained prior to calculate AHSN feature based on the eigenvalue decomposition of the Hessian matrix
•The orientation of surface normal vector is represented as two kinds of orientation in spherical coordination
85
86. •Two angular histograms are constructed
–The orientation θ histogram with n bins is formed
•Each bin covering 180/n degrees
•Each sample in the image block added to a histogram bin is weighted by its surface-ness saliency and normalized by total sum of surface-ness saliency
–The orientation φ is quantized into n bins
•Each bin covering 360/n degrees
•Each sample in the image block added to a histogram bin is weighted and normalized
–The dimension of feature descriptor is 2n
–The extracted AHSN feature is scale-invariant
86
87. 87
The extracted AHSN feature for a sphere (nodule model), left – reconstructed 3-D shape, center - orientation θ histogram, right - orientation φ histogram
88. 88
The extracted AHSN feature for a cylinder (vessel model), left – reconstructed 3-D shape, center - orientation θ histogram, right - orientation φ histogram
89. 89
The extracted AHSN feature for a curved surface (wall model), left – reconstructed 3-D shape, center - orientation θ histogram, right - orientation φ histogram
90. 90
The extracted AHSN feature for a pulmonary nodule, left – reconstructed 3-D shape, center - orientation θ histogram, right - orientation φ histogram
91. 91
The extracted AHSN feature for a pulmonary vessel, left – reconstructed 3-D shape, center - orientation θ histogram, right - orientation φ histogram
92. •Lung wall influence the detection accuracy
•For more accurate nodule detection, walls are eliminated from image blocks of nodule candidates
92
94. 94
Comparison of AHSN feature for a juxta-pleural nodule at before (1st row) and after (2nd row) wall elimination, left - reconstructed 3-D shape, center - orientation θ histogram, right - orientation φ histogram
95. 95
Comparison of AHSN feature for a solid nodule at before (1st row) and after (2nd row) wall elimination, left - reconstructed 3-D shape, center - orientation θ histogram, right - orientation φ histogram
96. •The extracted AHSN feature vectors are analyzed by SVM classifier
•SVM is a useful technique for data classification
•k-fold cross-validation is applied to evaluated the proposed classifier (k = 10)
96
97. Classfier
Accuracy
Sensitivity
Specificity
Before LWE
SVM-p
96.4%
98.4%
94.3%
SVM-r
97.8%
98.7%
96.9%
SVM-m
93.9%
95.9%
92.0%
After LWE
SVM-p
97.0%
97.9%
96.1%
SVM-r
97.8%
97.4%
98.2%
SVM-m
94.5%
94.6%
94.3%
97
The results of 10-fold cross validation on different kernel functions using SVM as a classier before and after wall elimination (LWE)
98. Descriptor
Accuracy
Sensitivity
Specificity
SVM-p
Gradient
95.1%
96.4%
93.8%
Hessian Matrix
97.0%
97.9%
96.1%
SVM-r
Gradient
96.1%
96.4%
95.9%
Hessian Matrix
97.8%
97.4%
98.2%
SVM-m
Gradient
92.8%
93.0%
92.6%
Hessian Matrix
94.5%
94.6%
94.3%
98
The results of 10-fold cross validation on with four different kernel functions based SVMs for the descriptors using gradient and Hessian matrix
100. 100
ROC curves of the SVM classifiers with respect to three different kernel
functions, SVM-r: radial basis function, SVM-p: polynomial function, and SVM-m:
Minkowski distance function; (a) p = 0:25 and (b) p = 1.
103. 103
(a) (b)
The result of pulmonary nodule detection: (a) 43rd slice, (b) 3-D representation, the detected nodules are indicated by a red color and the non-nodules are indicated by a white color
104. 104
AUC
Accuracy
Specificity
Sensitivity
FPs/scan
Nodule Candidates Detection
96.6%
51.25
20-80
0.921
76.6%
75.9%
88.3%
12.32
50-50
0.960
86.7%
86.4%
91.7%
6.99
80-20
0.967
89.6%
89.3%
90.9%
5.45
The results of CAD system using GP based classifier
105. 105
FROC curves of the GPC with respect to three training and testing datasets
107. 107
FROC curves of the proposed CAD system with respect to three different
kernel parameters of SVM-r classifiers
108. 108
The overall performance of CAD system for different parameters p of SVM-r classifiers
AUC
Accuracy
Specificity
Sensitivity
FPs/scan
Nodule Candidates Detection
97.9%
135.39
AHSN 180
0.9945
97.8%
98.2%
95.4%
2.43
AHSN 90
0.9923
97.5%
97.9%
95.2%
2.84
AHSN 72
0.9895
97.6%
97.7%
95.4%
3.11
109. 109
FROC curves of the proposed CAD system with respect to three different dimensions of AHSN features
110. CAD systems
Nodule size
FPs per case
Sensitivity
Suzuki et al.(2003)[26]
8 - 20 mm
16.1
80.3%
Rubin et al.(2005)[27]
>3 mm
3
76%
Dehmeshki et al.(2007)[28]
3 - 20 mm
14.6
90%
Suarez-Cuenca et al.(2009)[29]
4 - 27 mm
7.7
80%
Golosio et al.(2009)[30]
3 - 30 mm
4.0
79%
Ye et al.(2009)[31]
3 - 20 mm
8.2
90.2%
Sousa et al.(2010)[32]
3 - 40.93 mm
-
84.84%
Messay et al.(2010)[33]
3-30 mm
3
82.66%
Riccardi et al.(2011)[34]
>3 mm
6.5
71.%
Cascio et al.(2012)[35]
3-30 mm
6.1
97.66%
Genetic Programming
3-30 mm
5.45
90.9%
Hierarchical Block Analysis
3-30 mm
2.27
95.2%
Shape-based Feature
3-30 mm
2.43
95.4%
110
111. •Automated pulmonary nodule detection system is studied
•Pulmonary nodule detection CAD system is an effective solution for early detection of lung cancer
•The proposed systems are based on
–Genetic programming based classifier
–Hierarchical block-image analysis
–3-D shape-based feature descriptor
111
112. •The performance of the proposed CAD systems is evaluated on the LIDC database of NCI
•The GPC based system was shown to significantly reduce the false positives while maintaining a high sensitivity
–5.45 FPs/scan, 90.9% sensitivity
•The hierarchical block-image analysis based system has shown more accurate result with improved local object segmentation
–2.27 FPs/scan, 95.28% sensitivity
•Shape-based feature descriptor was applied the nodule detection CAD system that has shown higher accuracy and robustness than conventional descriptor
–2.43 FPs/scan, 95.4% sensitivity
•The proposed methods have significantly reduced the false positives in nodule candidates
112