This document describes a method for detecting pulmonary nodules in CT scans using genetic programming. It first segments the lung regions from CT images and extracts nodule candidates. Features are then extracted from the candidates. Genetic programming is used to classify candidates as nodules or non-nodules by optimizing combinations of features. The method was tested on a publicly available lung image database, achieving a true positive rate of over 90% and low false positive rate.

1. Computer-aided Detection of Pulmonary
Nodules using Genetic Programming
Wook-Jin Choi and Tae-Sun Choi

2. Contents
• Introduction
• Lung Segmentation based on 3D Approach
• Nodule Candidates Detection and Feature
Extraction
• Genetic Programming Based Classification
• Experimental Results
• Conclusions
• References
2

3. Introduction
• Pulmonary nodule detection is attractive applications of computer-aided
detection (CAD) because lung cancer is the leading cause of cancer
deaths.
• If lung cancer detected in early phase, the 3-year survival rate is more
than 80%.
• Recently, researchers have developed a number of CAD methods for lung
nodules to aid radiologists in identifying nodule candidates from CT
images.
• Current CT technology allows for near isotropic, submillimeter resolution
acquisition of the complete chest in a single breath hold.
• These thin-slice chest scans have become indispensable in thoracic
radiology, but have also substantially increased the data load for
radiologists.
• Automating the analysis of such data is, therefore, a necessity and this
has created a rapidly developing research area in medical imaging.
3

4. Related Works
• Template matching methods
– Genetic Algorithm Template Matching [10]
– 3D Template Matching [11]
• Model based methods
– Patient-specific models [5]
– Surface normal overlap model [7]
• Machine learning techniques
– Neural network [6]
– Fuzzy c-means clustering [9]
• Digital filtering
– Quantized convergence index filter [8]
– Iris filter [13]
• Statistical analysis [12]
4

5. Proposed Algorithm
Flow chart of Pulmonary nodule detection
5

6. Lung Segmentation based on 3D Approach
6

7. Lung Segmentation based on 3D Approach
• Select adaptive threshold value at every slice in the CT image
sequence using diagonal intensity histogram [4].
• The CT images are divided into background area(body) and
foreground area(air or lung) as shown below.
7
Original CT image and converted CT image with threshold

8. Lung Segmentation based on 3D Approach
8
• Segment lung region and remove the rim (outer part of the
body).
• Correct the contour of the lung volume (correct excluded wall
side nodule).
Extracted lung region using 3D connected component labeling and
contour corrected lung region (containing wall side nodule)

9. Nodule Candidates Detection and Feature Extraction
9

10. ROI Extraction
10
6-stepped ROI and extracted
nodule candidates
• Adaptive multiple
thresholding method.
– the traditional multiple
thresholding method makes
many steps of grey levels.
– We calculate the adaptive
threshold value using diagonal
histogram at every slice of
lung volume.
– This value is base threshold
value for multiple
thresholding.
– We make additional five
threshold values which are
base threshold + 50, -50, -100,
-150 and -200.

11. Nodule Candidates Detection
• We can remove the vessels and
noise in the lung volume using
rule based classifier.
• Vessel Removing
The vessel is classified by volume
elongation factor and compactness.
– volume is extremely bigger than nodule
– longer than nodule
– not compact object.
• Noise removing
– radius of ROI is smaller than 3mm or
bigger than 30mm.
• Remaining ROIs are nodule candidates
11
6-stepped ROI and extracted
nodule candidates

12. Feature Extraction
• 3D geometric features
– Volume
– elongation factor
– Compactness
– approximated radius.
• 2D pixel based features.
– Use median slice of nodule candidates (area of the median slice is the largest)
– To extract 2D texture feature, we normalize the image size of nodule
candidates.
– 3 types of nodule sizes and then extract the features.
• < 5mm : the size of image matrix is 8x8.
• 5mm ~10mm : the size of image matrix is 16x16.
• > 10 mm : the size of image matrix is 32x32
– extract 14 features from the image matrix.
• mean, variance, skewness, kurtosis, area, radius and 8 biggest eigenvalues.
12

13. Feature Extraction
13
Index Feature
1 Z position
2 Mean
3 Variance
4 Skewness
5 Kurtosis
6 Area
7 Radius
8 Perimeter
9 Compactness
10~17 Largest Eigenvalue 1~8
18 X centroid
19 Y centroid
20 Z centroid
21 Width
22 Height
23 Depth
24 Size

14. Genetic Programming Based Classification
14

15. Genetic Programming Based Classification
• Genetic Programming (GP)
– an evolutionary optimization technique [14].
• The basic structure of GP is very similar to
Genetic Algorithm(GA).
• The chromosome
– GA : variable (binary digit)
– GP : program (tree or graph)
15

16. Genetic Programming Based Classification
16
A function represented as a tree structure

17. Genetic Programming Based Classification
• Our goal of GP evolution is to reduce false positive
(FP) and increase true positive (TP).
• In the proposed scheme, an optimized classifier is
carried out using combination of features and random
constant values.
• GP optimally selects adequate features from all
extracted features and combines the selected features
with mathematical operators.
• The GP generates individual classifiers and those are
evaluated by fitness function.
• The result of GP can convert complex input features to
simple value.
17

18. Genetic Programming Based Classification
• GP chromosome
– The terminal set - The elements of feature vector extracted from
nodule candidate images and randomly generated constants with in
the range 0,1.
– The function set - Four standard arithmetic operator namely plus,
minus, multiply and division and 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.
18

19. Genetic Programming Based Classification
• 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).
TN FP
SPC FPR
    
– Area under the ROC curve (Az)
• ROC curve is plotted between TP and FP for different threshold values.
• Az is area under the ROC curve and a good measure of classifier performance in
different condition.
– Fitness Function
19
TP
TPR
TP FN


1 1
TN FP FP TN
 
f  TPR*SPC* Az

20. Genetic Programming Based Classification
Objective To evolve maximum fitness
Selection Generational
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
Genetic Programming parameter
20

21. Genetic Programming Based Classification
• Examples of GP
– minus(minus(P(21,:),exp(P(23,:))),minus(mypower(mylog(plus(times(P(14
,:),minus(P(23,:),mypower(mylog(plus(times(P(12,:),minus(P(11,:),mypow
er(P(13,:),P(13,:)))),P(22,:))),P(13,:)))),minus(P(20,:),cos(exp(P(7,:)))))),mypo
wer(exp(P(7,:)),P(7,:))),times(minus(minus(mypower(exp(P(23,:)),P(7,:)),P(
11,:)),times(exp(P(23,:)),P(12,:))),P(11,:))))
– minus(minus(plus(P(4,:),P(7,:)),sin(minus(P(7,:),mypower(P(24,:),plus(min
us(P(7,:),P(11,:)),mypower(P(15,:),P(7,:))))))),mypower(mypower(mypower
(P(24,:),exp(P(11,:))),minus(plus(P(10,:),mypower(plus(minus(P(13,:),sin(e
xp(P(12,:)))),mypower(P(24,:),plus(P(13,:),P(4,:)))),plus(plus(minus(0.35089
,0.35089),P(3,:)),P(7,:)))),P(11,:))),minus(plus(P(10,:),minus(P(4,:),plus(P(10,:
),P(7,:)))),mypower(minus(plus(P(4,:),exp(P(2,:))),minus(P(9,:),P(4,:))),P(11,:
)))))
21

22. Experimental Results
• Lung Image Database Consortium (LIDC) database [15]
– to evaluate the performance of the proposed method.
– 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 is separated into 84 cases, each containing around 100-400
Digital Imaging and Communication (DICOM) images and an XML data file
containing the physician annotations
• We applied our method to 32 scans consisting of 153 nodules
and 7528 slices. The pixel size in the database ranged from 0.65
to 0.75 mm and the reconstruction interval ranged.
• The half of dataset(16 scans) is used for training and another half
of dataset(another 16 scans) is used for testing the classifier.
22

23. Experimental Results
(a) (b)
The result of pulmonary nodule detection: (a) 43rd slice, (b) volume
rendering
23

24. Experimental Results
Data set TPR FPR Az
learn 93.33% 0.127 0.934
test 91.67% 0.138 0.897
all 92.31% 0.133 0.912
24
The results of GP based classifier

25. Experimental Results
ROC curves of GP based classifier with respect to three datasets
25

26. Conclusion
• We have proposed a novel pulmonary nodule detection algorithm in CT
images.
• Lung region is segmented using adaptive thresholding and voxel labeling
based method.
• Then nodule candidates are detected using adaptive multiple
thresholding and rule based classifier with 3D geometric features.
• Next, 3D and 2D features are extracted from the detected nodule
candidates.
• Finally, the extracted features are optimized and then classified into
nodule and non-nodule using GP.
• We applied proposed algorithm to the LIDC database of NCI.
• This method extremely reduced FP rate.
• The FPs per scan is only 6.5 with more than 90% sensitivity.
• The results show the superiority of the proposed method.
26

27. References
• [1] Ahmedin Jemal, Rebecca Siegel, ElizabethWard, Yongping Hao, Jiaquan Xu, and Michael J
Thun, “Cancerstatistics, 2009,” CA Cancer J Clin, vol. 59, no. 4, pp. 225–49, Jan 2009.
• [2] K-W Jung, Y-J Won, S Park, H-J Kong, J Sung, H-R Shin, E-Cl Park, and J S Lee, “Cancer
statistics in korea: incidence, mortality and survival in 2005,” J Korean Med Sci, vol. 24, no. 6,
pp. 995–1003, Dec 2009.
• [3] Qiang Li, “Recent progress in computer-aided diagnosis of lung nodules on thin-section
ct.,” Comput Med Imaging Graph, vol. 31, no. 4-5, pp. 248–257, 2007.
• [4] S G Armato, M L Giger, C J Moran, J T Blackburn, K Doi, and H MacMahon, “Computerized
detection of pulmonary nodules on ct scans,” Radiographics, vol. 19, no. 5, pp. 1303–11, Jan
1999.
• [5] M Brown, M McNitt-Gray, J Goldin, R Suh, J Sayre, and D Aberle, “Patient-specific models
for lung nodule detection and surveillance in ct images,” IEEE TMI, vol. 20, no. 12, pp. 1242 –
1250, Dec 2001.
• [6] K Suzuki, SG Armato III, F Li, S Sone, and K Doi, “Massive training artificial neural network
(mtann) for reduction of false positives in computerized detection of lung nodules in low-dose
computed tomography,” Medical physics, vol. 30, pp. 1602, 2003.
• [7] D Paik, C Beaulieu, G Rubin, B Acar, R Jeffrey, J Yee, J Dey, and S Napel, “Surface normal
overlap: a computer-aided detection algorithm with application to colonic polyps and lung
nodules in helical ct,” IEEE TMI, vol. 23, no. 6, pp. 661 – 675, Jun 2004.
• [8] Sumiaki Matsumoto, Harold L Kundel, James C Gee, Warren B Gefter, and Hiroto Hatabu,
“Pulmonary nodule detection in ct images with quantized convergence index filter.,” Med
Image Anal, vol. 10, no. 3, pp. 343–352, Jun 2006.
27

28. References
• [9] N Memarian, J Alirezaie, and P Babyn, “Computerized detection of lung nodules with an
enhanced false positive reduction scheme,” IEEE ICIP 2006, pp. 1921 –1924, Sep 2006.
• [10] Jamshid Dehmeshki, Xujiong Ye, Xinyu Lin, Manlio Valdivieso, and Hamdan Amin,
“Automated detection of lung nodules in ct images using shape-based genetic algorithm.,”
Comput Med Imaging Graph, vol. 31, no. 6, pp. 408–417, Sep 2007.
• [11] Onur Osman, Serhat Ozekes, and Osman N Ucan, “Lung nodule diagnosis using 3d
template matching.,” Comput Biol Med, vol. 37, no. 8, pp. 1167–1172, Aug 2007.
• [12] A El-Baz, G Gimel’farb, R Falk, and M Abo El-Ghar, “Automatic analysis of 3d low dose ct
images for early diagnosis of lung cancer,” Pattern Recognition, vol. 42, no. 6, pp. 1041–1051,
Jan 2009.
• [13] JJ Su´arez-Cuenca, PG Tahoces, M Souto, MJ Lado, M Remy-Jardin, J Remy, and J Jos´e
Vidal, “Application of the iris filter for automatic detection of pulmonary nodules on
computed tomography images,” Computers in Biology and Medicine, 2009.
• [14] J Koza, “Genetic programming: On the programming of computers by means of natural
selection,” The MIT Press, Jan 1992.
• [15] S G Armato, G McLennan, M F McNitt-Gray, C R Meyer, D Yankelevitz, D R Aberle, C I
Henschke, E A Hoffman, E A Kazerooni, H MacMahon, A P Reeves, B Y Croft, L P Clarke, and
Lung Image Database Consortium Research Group, “Lung image database consortium:
developing a resource for the medical imaging research community.,” Radiology, vol. 232, no.
3, pp. 739–748, Sep 2004.
28

29. Thank You
29