Computer aided detection of pulmonary nodules using genetic programming

Embed Size (px)

DESCRIPTION

Computer aided detection of pulmonary nodules using genetic programming

Text of Computer aided detection of pulmonary nodules using genetic programming

  • 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,:)))))),m