使用基因演算法變數篩選與SVM分類器於PET/CT上孤立肺結節之診斷

全文

(1) . 國立高雄大學資訊工程學系(研究所) 碩士論文 . 使用基因演算法變數篩選與 SVM 分類器於 PET/CT 上孤立肺結節之診斷 Diagnosis of solitary pulmonary nodule in PET/CT using GA for feature extraction and SVM classifiers . 研究生：周書賢撰指導教授：殷堂凱博士. 中華民國一百零二年七月 i . .

(2) 使用基因演算法變數篩選與 SVM 分類器於 PET/CT 上孤立肺結節之診斷指導教授：殷堂凱博士（教授）國立高雄大學資訊工程學系學生：周書賢國立高雄大學資訊工程學系碩士班摘要醫學上的電腦斷層掃描技術(Computed Tomography，CT)方便醫師以肉眼判斷病患身體內部的異常部份，以診斷出腫瘤的形狀與位置。現常與正電子發射掃描計算機斷層掃描(Positron Emission Tomography，PET)並用，以求更快速、準確找出腫瘤。對於肺部腫瘤，診斷主要方法是在 CT 上針對肺結節(lung nodule)的形狀、大小與位置做判斷，也以 PET 上的標準攝取值(Standardized uptake value，SUV 值)數值大小判斷。此篇論文針對 68 位病患(40 位癌症患者，28 位為非癌症之良性結節患者)，察看診斷報告擷取 65 顆孤立結節，作門檻值與取輪廓作完處理，再進行良性與惡性分類，再以統計與紋理的特徵(Texture Features)提取變數，輸入基因演算法(Genetic Algorithm)作變數篩選，並用支持向量機(Support Vector Machine，SVM)分類，由電腦自行訓練與辨識良性結節與孤立的惡性腫瘤，找出最適當的重要特徵變數。使用 5-fold cross validation 與 5 次重覆實驗，GLCM 四個方向所有的變數篩選出 22 個變數他們的敏感度是 72.41%特異度 72.22% 準確率 72.30%。若使用十四個變數敏感度特異度準確率分別是 79.31%，72.22%和 75.38%。若篩選五個變數，它們的敏感度是 79.31%，特異度是 80.56%，準確率是 80.00%。這篇論文顯示變數組合的紋理特徵有助於提高分類良性與惡性腫瘤的準確率。關鍵字：PET/CT、標準攝取值(SUV)、支持向量機、基因演算法、紋理特徵. ii .

(3) Diagnosis of solitary pulmonary nodule in PET/CT using GA for feature extraction and SVM classifiers. Advisor: Dr. Tang-kai Yin Department of Computer Science and Information Engineering National University of Kaohsiung Student: Shu-hsien Chou Department of Computer Science and Information Engineering National University of Kaohsiung ABSTRACT The Computed Tomography (CT) scan technique in medical science enables physicians to diagnose the abnormal parts within the body of the patients with unaided eyes. It is a diagnostic tool used to detect the shape and position of the tumors. It is now jointly used with Positron Emission Tomography (PET) scan for a fast and accurate way to find out the tumors. The diagnostic method for lung tumors is to determine the shape, size and position of pulmonary nodules from the image of CT scan, taking into account the Standardized uptake value (SUV) from PET. This study was done on 68 patients (40 cancer patients, 28 non-cancer patients with benign nodules). It investigated the diagnostic report of 65 solitary nodules, and classified them into benign or malignant tumors after treating for threshold value and contouring. The tumors were subsequently computed according to their texture features to find out all the variables, applying with Genetic Algorithm (GA) for variable screening and using Support Vector Machine (SVM) for their classification. The most appropriate variable characteristics for the determination of benign and malignant tumors were found out using computer automatic training. The final result of this study was obtained from the 5-times-repeating experiment in the 5-fold cross-validation. When 22 variables were screened out of all the variables in the four directions in GLCM, their sensitivity was found to be 72.41%, specificity 72.22% and accuracy 72.30%. If 14 variables were used instead, the sensitivity, specificity and accuracy were 79.31%, 72.22% and 75.38% respectively. When 5 variables were screened, they showed a sensitivity of 79.31%, specificity 80.56% and accuracy 80.00%. This study indicates texture features from the combination of variables are useful for enhancing the accuracy for classifying benign and malignant tumors. Keywords: Positron emission tomography combined computed tomography (PET/CT), Standardized Uptake Value (SUV), SVM, Genetic Algorithms, Texture Features iii .

(4) Content . Chapter 1 Introduction .................................................................................... 1 1.1 Related Works ......................................................................................... 2 1.2 Motivation ............................................................................................... 7 1.2.1 Texture analysis ......................................................................... 7 1.2.2 PET/CT study ............................................................................ 8 1.3 Thesis Layout .......................................................................................... 9 Chapter 2 Fundamentals of Feature Extraction and Machine Learning. 10 2.1 Calculation of Standardized Uptake Value ........................................... 10 2.2 Feature Extraction ................................................................................. 12 2.3 Genetic Algorithms ............................................................................... 20 2.4 Support Vector Machine ........................................................................ 21 Chapter 3 Diagnosis of Solitary Pulmonary Nodule in PET/CT Using GA for Feature Extraction and SVM Classifiers ............................ 23 3.1 Procedure ............................................................................................... 23 3.2 Materials ................................................................................................ 24 3.3 Merge PET/CT ...................................................................................... 25 3.4 Image Segmentation .............................................................................. 26 3.5 Preprocessing ........................................................................................ 27 3.5.1Thresholding ................................................................................ 27 3.5.2 Contour ...................................................................................... 28 3.6 Variable Computation............................................................................ 30 3.7 Variable Selection .................................................................................. 31 3.8 Cross-validation .................................................................................... 33 Chapter 4 Experiment Results...................................................................... 35 Chapter 5 Conclusion .................................................................................... 43 References .......................................................................................................... 45 Appendix A ........................................................................................................ 50 Appendix B ........................................................................................................ 54 . . iv .

(5) List of Figures . Figure 2- 1 Geometrical relationships of GLCM measurements made for 4 distances d. ................................................................................. 13 Figure 2- 2 Illustration of texture calculation using GLCM technique. (a) A matrix representation of a 5 × 5 pixel image with three grey values; (b) the GLCM P (i, j) for d = [1, 1]. .............................. 14 Figure 3- 1 The flow chart of the experiment. ............................................... 24 Figure 3- 2 The axial, coronal and sagittal cross sections of a pulmonary nodule from top to bottom. (a) CT image; (b) PET image; (c) merged PET/CT image. The highlighted regions as shown in the figure are the pulmonary nodules with higher SUV. .................. 26 Figure 3- 3 Thresholded image of a pulmonary nodule. (a) the original image; (b) the thresholded image using a gray-value threshold of 0.5 on image (a). ................................................................................... 28 Figure 3- 4 Contouring the image of a pulmonary nodule. (a) The original image; (b)-(e) a red line was iterated to form a circle around the tumor according to its internal energy term, until (f) a dividing line was caught around the tumor boundary; (g) a contoured image of a pulmonary nodule. ................................................... 29 Figure 3- 5 Manual dividing line set up to obtain a contoured image. (a) The original image; (b)-(c) a red line was iterated to circle the tumor according to its internal energy term; (d) the dividing line exceeded the image boundary; (e) a manual dividing line was set up to surround the desired part of the tumor; (f) a contoured image of a pulmonary nodule. ................................................... 30 Figure 3- 6 Images of a segmented malignant pulmonary nodule with its description. Starting from left to right is the original CT image, thresholded and contoured image, and the PET image. ............. 31 Figure 3- 7 Maximum and average fitness. ................................................... 32 Figure 3- 8 The frequency of bit 1 at (a) generation = 0, and (b) generation = 10 resulting from GA development. .......................................... 33 . . v .

(6) List of Tables Table 1- 1 List of the HU of the several substances .......................................... 4 Table 4- 1 Sensitivity, specificity and accuracy, and their mean and standard deviation of classifications in each direction to GA ..................... 36 Table 4- 2 The number of occurrences of each variable ................................. 38 Table 4- 3 The numbers of those higher frequency variances ........................ 39 Table 4- 4 Sensitivity, Specificity and Accuracy of three sets of selected variables using SVM only ............................................................. 41 . vi .

(7) Chapter 1. Introduction. . Malignant tumors, including trachea, bronchus and lung cancer, which account for about one-fifth, the largest proportion of the number of cancer deaths, have become the first of the top ten leading causes of death in Taiwan in the last thirty consecutive years (行政院衛生署, 2013). Due to the advancing medical technology, there are many ways for physicians to detect cancer today. The physicians often use computed tomography (CT), to diagnose abnormal parts of the patients by using X-ray tubes and the three-dimensional image displayed in the computer. Computed tomography helps physicians determine and diagnose the shape, size and position of the tumor inside the patient’s body. Physicians also use state-of-the-art clinical examination imaging technology: positron emission tomography (PET) in the field of nuclear medicine as a tool to classify the pulmonary nodules as either benign or malignant more accurately. The shape, size and position that are used for the clinical judgment in the diagnosis of pulmonary nodules are not only observed on CT, but also by standardized uptake value (SUV) in PET. SUV is a semi-quantitative variable commonly used to analyze images of tumors. injected. with. a. radiopharmaceutical. called. FDG. ( fluoro-2-deoxy-d-glucose ). The cut-off value for malignant lesions was about 2.5 which had been shown as an indicator for tumor classification (Masa-ah et al., 2010). However, not all patients with high SUV in PET suffer from cancer, 1 .

(8) so one of the subjects of our study is to focus on a variety of numerical SUVs in the merging of CT and PET images. Benign tumors in lung could be classified into bronchial adenomas, bronchial fibromas, hemangiomas, etc., while malignant tumors could be classified into small cell lung cancer, non-small cell lung cancer, adenocarcinoma, large cell carcinoma etc. There have been a lot of researches done on the CT of pulmonary nodules, including how to cut the lungs, capture the nodules automatically, analyze their features, delineate the nodules, etc. Using the images to determine the tumors as either benign or malignant does not necessarily mean the fact, however, tumors biopsy still need to be done on patient eventually. In order to determine whether the nodules in high-SUV patients are benign or malignant, we obtained the patients’ image data from DICOM (Digital Imaging and Communications in Medicine) file provided by physicians. We merged the CT and PET images, segmented the lung section, selected suspicious nodules, calculated the features, and finally applied genetic algorithm and SVM analysis to find the most useful factors that can be used to detect cancer. By comparing and analyzing many high-SUV patients with benign pulmonary nodules, we determine which image features are more significant in the diagnosis of pulmonary tumors.. 1.1 Related Works Physicians are using more precise machines and equipment in medical 2 .

(9) diagnostic imaging due to the development in medical technology. They use shooting rays from different directions to diagnose the tumors in human body organs. There are a variety of instruments that form the common image sources, including ultrasound, magnetic resonance imaging, single-photon emission computed tomography, X-ray, computed tomography, positron emission tomography, etc. As a result of the increasing capacity of storage devices for image information, DICOM was invented to effectively classify and manage the image data according to their characteristics. In the field of medical imaging, many studies concerning the CT image processing methods have been done. The CT image can be analyzed according to its features extraction and classified after computation. Each pixel on the CT image represents the attenuation coefficient of the analyzing substance on that point; its dimension in HU (Hounsfield Unit) can be calculated by the following formula: C T N um ber =. . μ − μw ×K μw. (1-1). μ and μw represent the attenuation coefficient of the substance and. water respectively, where the value of K is usually 4096. The attenuation coefficient of a substance is related to its density, substances with higher density have greater CT numbers. CT number in air and bone ranges from -2048 to +2048（曾林維，2006）. HU is a useful practical unit. For a substance with linear attenuation coefficient μ, the HU scale (indicative value) can be obtained from CT 3 .

(10) numbers by:. HU =. μ − μw × 1000 μw − μa. (1-2). Distilled water is defined as 0 HU, while air is defined as -1000 HU. Others such as lung is -500, muscle is 40, and soft tissue is 100 to 300. The HU value of bone is relatively high, about 700 to 3000, so the bone mostly displays white image.. Table 1- 1 List of the HU of the several substances Substance. HU. Air. -1000. Lung. -500. Fat. -100 to -50. Water. 0. Kidney. 30. Blood. +30 to +45. Muscle. +10 to +40. Liver. +40 to +60. Soft Tissue. +100 to +300. Bone. +700 to +3000. For automatic detection of pulmonary nodules in CT, a series of stages, namely thorax extraction, lung extraction, lung reconstruction, structures 4 .

(11) extraction, tubular structures elimination and false positive reduction were used. 3D visualization of the structures was the main approach taken to correct identification of pulmonary nodules in automatic detection (da Silva Sousaa, et al., 2010). Cutting out the contour of the image from 3-D X-ray CT data, which was the organization of a stack of 2-D transverse slices of pulmonary nodules for quantization, it was found that air had a mean intensity of approximately -1000 Hounsfield units (HU) in the CT image, whereas lung tissue was in the range between -910 HU to -500 HU. By using this method, lung extraction and left and right lung separation can be done (Hu and Reinhardt, 2001). Shapes and fractals are used as the indicators in diagnosing pulmonary nodules. A radiologist can distinguish benign masses or malignant breast tumors by studying their shape characteristics, which are represented by the turning angle functions of their contours. The turning angle function could be used as a signature to represent a contour’s shape characteristics in mammograms, so it could be used to classify the contours. It is nearly monotonically increasing in the contours of benign masses, while it is decreasing and increasing in the contours of malignant tumors (Rangayyan et al., 2006). In addition, a study from Atam P. et al. provided a collective numbers of Gray-level Co-occurrence Matrix (GLCM)-related formulas and methods in finding the texture of CT image, and their distributions support a lot in analyzing and comparing the textures to classify the malignant and benign pulmonary nodules (Dhawan A.P. et al., 1996, Albregtsen, F., 2008). 5 .

(12) Compared to many analyzing studies concerning pulmonary nodules on CT imaging, fewer studies were done together with PET. Examining PET image is a new screening way which is very helpful in medical diagnosis, and there have also been a lot of medical applications done on it. For instance, the study of PET in detecting extraosseous myeloma had shown that radiation played an important factor in the initial workup and follow-up evaluation (Hall, M.N., 2010). . In recent years, joint studies of CT and PET have been implemented to. detect the region of the nodules using the foreground and background mean ratio independently in oncology. Cases when nearby and similar nodules were merged into one by a split-up post-processing step were also dealt with in PET/CT studies (Zsoter, N., 2012). Research indicated that max SUV cutoff larger than 2.5 is a very fine indicator to distinguish between the benign and malignant for nodules larger than 1 cm in size, but it does not seem to be completely useful for nodules less than 1 cm (Khalaf, M. et al., 2008). Studies showed exceptions that using a maximum SUV cutoff of 2.5 to determine nodules on PET image was not feasible as many nodules were benign regardless of their high SUVs. Their high SUVs could be due to inflammation after the surgery or other reasons. Therefore, the study of high-SUV nodules from diagnostic reports is to analyze the differences of the factors that affect the quality of diagnosis. Filtering out many high degree characteristics of features for classification is an important research direction in our study. 6 .

(13) 1.2 Motivation McNitt-Gray studied the nodules classification by inputting solitary nodules to a linear discriminate classifier to predict the pulmonary nodules as benign or malignant based on quantitative measures extracted from high-resolution CT image. The quantization of nodule used co-occurrence matrices which were formed by different combinations of gray level quantization and distance between pixels and angles. There are many new developments in further studies.. 1.2.1 Texture analysis 曾林維 analyzed the texture changes of acute middle cerebral artery (MCA) ischemic stroke on CT images based on the quantification of computed textures. By removing useless information and delineating the boundary between MCA territories, the following texture features including entropy, energy, contrast, homogeneity and low density number were calculated. The experimental results showed that by using the computed texture quantification, the accuracy for system using back propagation neural classifier and self-organizing map classifier were 0.95 and 0.74 respectively, much better than the clinician's stroke diagnostic trial, thus providing more visional diagnostic information. 鄭煥勳 acquired 82 solitary pulmonary nodules on CT images and computed texture features such as entropy, energy, contrast, and homogeneity, in addition to the area, diameter, circularity, etc. by using co-occurrence matrix for parameters analysis. The area under Receiver Operating Characteristic (ROC) 7 .

(14) curve was evaluated, and then integrated with back propagation neural network. The accuracy of identification for benign and malignant was 0.791, indicating that the texture features will change at different times throughout the development. This study merged CT with PET to determine the benign and malignant nodules and it also showed the importance of using texture parameter.. 1.2.2 PET/CT study Potesil, V. et al. used a new method to delineate tumor boundaries by integrating CT classification with PET classification. Information from PET and CT images was established by the different contrast between them to derive the final delineation. It was proposed that PET delineation result was improved and segmentation could be achieved with minimal user interaction. This method, however, failed to apply to all the nodules because some of them do not possess complete cutting shape and contain high FDG uptake. Han, D. et al. made use of the Markov Random Field (MRF) based image segmentation in his study. MRF is a form of undirected graphical model which is more appropriate to produce discriminative classifiers, penalizing the segmentation difference between PET and CT images for PET/CT co-segmentation. By taking the individual domain ratios of CT and PET, and capturing the neighboring spaces, the SUV of the suspected nodules relative to PET on CT could be measured. The purpose of our study is to use image processing techniques and computed visional textures quantitative analysis to construct an expected 8 .

(15) computed aided diagnosis (CAD) system and analyze the information provided by the features of pulmonary nodules in CT and PET to increase sensitivity, specificity and accuracy of the pulmonary nodules of PET/CT for pulmonary nodules evaluation.. 1.3 Thesis Layout In chapter 2, we introduce the fundamentals of feature extraction and machine learning. Chapter 3 shows the diagnosis of solitary pulmonary nodule in PET/CT using GA for feature extraction and SVM classifiers. In chapter 4, we demonstrate the experimental results of our study. Chapter 5 presents how our study benefits physicians in cancer diagnosis and its further studies.. 9 .

(16) Chapter 2. Fundamentals of Feature Extraction and Machine Learning. . 2.1 Calculation of Standardized Uptake Value Positron emission tomography (PET) scan appeared in the 1970s, mainly in academics at the beginning. It has a lot of applications in cancer, heart disease and neurological diagnosis of mental illness and it is a breakthrough to become a gradually important clinical diagnostic tool. Fluoro-2-deoxy-d-glucose (FDG) is a radiopharmaceutical used with PET scanner to provide diagnoses of various medical conditions. A PET scanner can form images of the distribution of FDG around the whole body which are mainly used for cancer checks. After intravenous injection of 5 to 10 millicuries (mCi) of FDG into the patient’s body, the subject should lie down to rest for 45 minutes for the accumulation of FDG in the tumor. As for normal tissues it can be fully removed and excreted by the kidneys and bladder. FDG-PET has turned into an important imaging modality. In the thorax, FDG-PET could distinguish benign pulmonary lesions and stage lung cancer from malignant ones. It is often more accurate than conventional imaging studies, and has been shown to be useful and cost-effective for nodules evaluation (Marom, E.M., et al. 2000). Since PET imaging only provides metabolic functional imaging, interpretation may need a computed tomography (CT) or magnetic resonance imaging (MRI) to obtain the related position of the organs. Visual interpretation 10 .

(17) is the most commonly used method, but the anatomical structure image fusion method with PET functional imaging can show more accurately the position of the foci and the extent of metabolic abnormalities. PET is used in the identification of benign and malignant lesions, pre-treatment staging, restaging treatment, treatment assessment and tracking on relapse after treatment etc. Under some conditions, the use of PET is quite effective. For example, PET can be used to diagnose a relapsed tumor, where anatomical imaging cannot diagnose exactly. Anatomical imaging cannot precisely diagnose the residual mass after treatment and confirm whether a tumor is benign or malignant. During the preoperative stage of lung cancer, CT cannot examine exactly either. Three steps were taken in order to calculate SUV: conversion of pixel value to activity concentration by multiplying the pixel value by rescale slope and 0.027027027; findings of decay calibration factor by taking the values of Total Dose (0018, 1074), Series Date (0008, 0021), Series Time (0008, 0031), Radiopharmaceutical Start Time (0018, 1072), and Radiopharmaceutical Half life (0018, 1075); and lastly inputting the patient body weight to the equations for the calculations of SUV (0010, 1010) (陳玥汝，2011).. SUV =. Activity Concetrati on ⎛ Injected Dose ⎞ ⎟⎟ ⎜⎜ ⎝ Body Weight ⎠. 11 . (2-1).

(18) SUV =. ( Pixel. Value × Rescale Slope ) × Body Weight Time Difference × Injected Dose 2 Isotope Hale −life. (2-2). . 2.2 Feature Extraction One way to describe quantified area is according to its texture feature. Feature extraction is a computed visional and image processing concept, which has so far no universal characteristics and precise definition. The image textures can provide intuitively smooth and rough measurements and consistency properties for measurements of parameters leading to direct image understanding. The image textures can be described by three main methods: statistical, structural and spectral methods. Texture features such as smooth, rough, and others are the representations of statistical methods. Structural image processing provides the unit technology arrangement. Spectral analysis is based on the nature of the spectrum, detecting an image's overall periodicity by recognizing high-energy peak in the spectrum. This study calculated the area of CT image, the maximum value of SUV on PET, and the average means of CT and PET images, and then applied Gray-level Co-occurrence Matrix (GLCM) technique to measure the image textures. GLCM created from a gray-scale image is defined as the distribution of the co-occurring values over that image. It is an approach to extract second or higher order statistical texture features. It calculates how often two adjacent 12 .

(19) pixels, one with grayscale intensity i and another with intensity j, occur when they are separated by a pixel distance (Albregtsen, F., 2008). As shown in Figure 2-1, the GLCM was formed by using the distance vectors: d [0 1], d [-1 1], d [-1 0] and d [-1 -1] as a set of offsets sweeping through 0, 45, 90 and 135 degrees. Figure 2-2 shows a matrix representation of a 5 × 5 pixel image with three grey values with its GLCM P (i, j) for d = [1, 1], where i and j are the pair pixel originated from different locations in the same grayscale image. Texture features can be computed by normalizing the GLCM. The normalized GLCM as the second-order histogram H(yq, yr, d) provides the probability of occurrence of a pair of gray values yq and yr separated by a distance vector d (Dhawan, A.P., 2011). 90° d[-1, 0] 135°. 45°. d[-1, -1]. d[-1, 1] 0° d[0, 1]. . Figure 2- 1 Geometrical relationships of GLCM measurements made for 4 distances d.. 13 .

(20) 2 0 0 1 2. 2 2 1 2 1. 2 2 1 2 0. . 0 1 2 0 1. 1 1 0 1 1 . (a). 0 2 1 0. 3 1 4 1. 1 0 3 2. 0 1 2. i. j. . (b). Figure 2- 2 Illustration of texture calculation using GLCM technique. (a) A matrix representation of a 5 × 5 pixel image with three grey values; (b) the GLCM P (i, j) for d = [1, 1]. . (1) Entropy (SH) of H(yq ,yr, d) yt. SH = − ∑. :. yt. ∑ H ( y , y , d) log. yq = y1 yr = y1. q. r. 10. [ H ( yq , yr , d )]. (2-3). Entropy can measure the texture non-uniformity; low values of entropy indicate greater structural variation.. (2) Angular second moment (ASMH) of H(yq ,yr, d). ASM H =. yt. yt. ∑ ∑ [H ( y. yq = y1 yr = y1. q. :. , y r , d )]2. (2-4). ASM is the degree of homogeneity among textures, and is also representative of the energy in the image. Low values of ASM indicate finer textures. 14 .

(21) (3) Contrast ( μ H ) of H(yq ,yr, d): m. μH = m. yt. yt. ∑ ∑. y r = y1 y r = y1. ∂ ( yq , yr ) H ( yq , yr , d ). (2-5). 2 where ∂ ( y q , y r ) = ( y q − y r ). Contrast values indicate the pixel intensity.. (4) Inverse difference moment (IDMH) of H(yq ,yr, d):. IDM H =. yt. yt. H ( yq , yr , d). ∑ ∑ 1 + ∂( y. yq = y1 yr = y1. q. , yr ). (2-6). IDM provides a measure of the local homogeneity among the textures instead of all, and the formula from MATLAB is different in the denominator.. (5) Inverse difference moment (IDMH) of H(yq ,yr, d) (in MATLAB):. IDM H =. (6) Correlation (CorH) of. CorH =. q. yt. ∑ ∑. y q = y1 y r = y1. yt. yt. ∑ ∑(y. r. y q = y1 y r = y1. q. 1 + yq − yr. − μ yq )( yr − μ yr ) H ( yq , yr , d) 15 . . H ( y q , y r , d) (2-7). H(yq ,yr, d) (In MATLAB):. 1. σy σy. yt. (2-8).

(22) In addition to the formulas as shown in (Albregtsen, F., 2008), we also applied the statistical texture formulas as described in (Dhawan A.P., 2011). (7) Correlation: G −1 G −1t. CORRELATION = ∑∑ i =0. where H m ( y q , d ) =. yr. ∑ H(y. y r = y1. q. {i × j}× P(i, j ) − {μ x × μ y } σ x ×σ y. j =0. , y r , d ) and H m ( y r , d ) =. (2-9) yr. ∑ H(y. y q = y1. q. , yr , d ). The correlation is greater for similar elements within the second-order histogram.. (8) Variance: G −1 G −1t. Variance = ∑ ∑ (i − μ ) 2 P (i, j ) i =0. j =0. (2-10). Variance is a measurement of heterogeneous area of image, showing the extent of deviation from the mean of the sample.. (9) Sum Entropy (SENT): 2G −2. SENT = − ∑ Px + y (i ) log( Px + y (i )) i =0. (2-11). G −1 G −1t. Px + y ( k ) = ∑ ∑ P (i , j ) i =0. (2-12). j =0. i + j = k for k = 0,1,..., 2(G − 1). Sum Entropy is the entropy of a new matrix produced by the summation of 16 .

(23) X-axis or Y-axis of GLCM, in which larger or smaller entropy can be obtained. The gray value in the new matrix changes more intensely to generate more information.. (10) Difference Entropy (DENT): G −1. DENT = − ∑ Px − y (i ) log( Px − y (i )) i= 0. (2-13). G −1 G −1t. Px − y ( k ) = ∑ ∑ P (i , j ) i =0. (2-14). j =0. | i − j |= k for k = 0,1,..., (G − 1). Difference Entropy is calculated based on the histogram of region difference images. The minimum value for DENT means most similar regions.. (11) Cluster Shade (SHADE): G −1 G −1. SHADE = ∑∑ {i + j − μ x − μ y } × P(i, j )) 3. i = 0 j =0. G −1 G −1t. μ x = ∑ i ∑ P (i , j ) i =0. j =0. G −1. G −1t. i =0. j =0. μ y = ∑ j ∑ P (i , j ). The cluster shade value is weighted for the neighborhood size.. (12) Cluster Prominence (PROM): 17 . (2-15).

(24) G −1 G −1. PROM = ∑∑ {i + j − μ x − μ y } × P (i, j )) 4. i =0 j =0. (2-16). Cluster prominence and cluster shade are quite similar in which cluster shade is a cubic equation, whereas cluster prominence is a quadratic equation.. The following equations are based on (Haralick, R.M. et al, 1973). (13) Sum Average (SA): 2 Ng. SA = ∑ ipx + y (i). (2-17). i =2. N g =Number of distinct gray levels in the quantized image. (14) Sum Variance (SV): 2Ng. SV = ∑ (i − SENT ) 2 Px + y (i ) i=2. (2-18). SENT is Sum Entropy which can be calculated as shown by equation in 2-11.. (15) Difference Variance (DV): DV=variance of. px− y. (2-19). (16) Information Measure of Correlation 1: IMC (1) =. HXY − HXY 1 max{HX , HY }. 18 . (2-20).

(25) . . HXY = −∑∑ p(i, j ) log( p(i, j )) i. (2-21). j. . HXY 1 = −∑∑ p(i, j ) log{ px (i ) p y ( j )} i. (2-22). j. (17) Information Measure of Correlation 2: IMC (2) = (1 − exp[−2.0( HXY 2 − HXY )]). 1. 2. (2-23). where HX and HY are entropies of Px and Py, and . HXY 2 = −∑∑ px (i ) p y ( j ) log{ px (i ) p y ( j )} i. j. (2-24). . The following equations are based on (Clausi D.A., 2002). (18) Inverse difference normalized (INN): IN N =. C ij. G. ∑. i , j =1. 1+ i − j. 2. /G2. (2-25). (19) Inverse difference moment normalized (IDN): IDN =. Cij. G. ∑ 1 + (i − j ). i , j =1. 2. / G2. The following equations are based on (Soh, L.K. and Tsatsoulis, C.) 19 . (2-26).

(26) (20) Autocorrelation (AC): AC = ∑∑ (ij ) p(i, j ) i. j. (2-27). (21) Dissimilarity: Dissimilarity = ∑∑ i − j ⋅ p(i, j ) i. j. (2-28). (22) Maximum probability (MP): MP = MAX (i, j ) i, j. (2-29). 2.3 Genetic Algorithms The Genetic Algorithm (GA) concept applied in our study was based on Mitchell’s 1999 book An Introduction to Genetic Algorithm. Through recombination, each locus having two possible alleles (0 and 1) in the chromosome is replaced with loci in other chromosomes. Each chromosome was assigned with an average fitness value. The populations of chromosomes were processed by GA to screen out chromosomes (candidates) with excellent fitness values for the next generation. We set the numbers of generation to 10 to select candidates for SVM classification. GA operation is simplified below: an initialized population of binary strings was evaluated and selected for the next population. Mutation and crossover operators were applied within this population to generate new strings that will be evaluated and added to the population. The process is repeated until a 20 .

(27) stopping criterion is met, that is, the population for the next generation is filled.. 2.4 Support Vector Machine Support vector machine (SVM) is a supervised learning model, as well as a binary linear classifier that can be used for linear and non-linear classification. It performs data clustering, achieving higher accuracy than traditional methods. Groups of data (vectors) are mapped to form two support hyperplanes, in which a classification hyperplane is constructed maximizing the distance (margin) between them. The idea of SVM is to find this constructed classification hyperplane to classify the data. The formula of classification hyperplane is defined as wT x + b = y. (2-29). And two support hyperplanes using a constant are written as: wT x + b = 1. (2-30). w T x + b = −1. (2-31). If the distance between the two formulas is given by. 2 w. The formula can be rewritten as yi ((wT xi ) + b) ≥ 1. (2-32). To optimize the distance between two support hyperplanes, the maximum 2 needs to be and their mean minimize needs to be w w 2. 21 .

(28) The Lagrangian can be written as. L(w, b,α ) =. 1 2 N w − ∑αi ⎡⎣ yi (wT xi + b) −1⎤⎦ 2 i =1. (2-33). Making use of the Lagrangian Duality theory, and optimizing L(w,b,α) with respect to w, b and α, this Lagrangian can be solved: 1 Minimize : LD = ∑ α i − ∑ α iα j yi y j xiT x j 2 ij i =1 Subject to. ∑α y i. i. =0. i. 22 . α i ≥ 0∀i. (2-34).

(29) Chapter 3. Diagnosis of Solitary Pulmonary Nodule in. PET/CT Using GA for Feature Extraction and SVM Classifiers 3.1 Procedure As shown in Figure 3-1, the experimental process started with obtaining the CT and PET images from the DICOM file, then handled the CT and PET images separately, and lastly combined them into PET/CT images. The images of suspicious nodules were diagnosed by physicians to divide them into benign and malignant nodules. Image pre-processing was performed which implemented the thresholding method prior to the delineation for the second treatment. After processing the images, the variables of the treated image area were computed before applying with GA and SVM for subsequent screening. Lastly, we identified which combinations of variables that result in malignant tumors.. 23 .

(30) DICOM. . CT. . PET. Image segmentation. . Malignant. Benign. . Pre-processing. . Variable Computation. . GA applied for feature extraction and SVM applied for classification. . Output. . Figure 3- 1 The flow chart of the experiment.. . 3.2 Materials In this study, we took the PET/CT images of 68 patients together with their medical diagnostic report from Digital Imaging and Communications in Medicine (DICOM) file provided by nuclear medicine physicians from National Cheng Kung University (NCKU) College of Medicine in Tainan City, Taiwan. 28 out of the 68 patients were recognized as having a benign nodule; the remaining 40 patients were recognized as having a malignant tumor.. 24 .

(31) 3.3 Merge PET/CT CT and PET images from the DICOM file were merged to become PET/CT image in this study. Since the CT images of blood vessels, kidneys, liver and other organs produced by X-ray are grayscale images, they need to be whitened until a clearer display of various human organs can be constructed. The purpose of using PET imaging is to identify abnormal metabolism of the cancerous organs and tissues, but PET imaging is not clearer than CT imaging. There are differences between these two different image resolutions; the CT images have a dimension of 512 × 512, while the dimension of PET images is of 168 × 168 only. Either CT or PET alone has specific benefits and shortcomings, but by merging CT and PET, physicians can diagnose and localize the tumors more accurately. In order to merge these two different images, PET image was up-sampled by 3.04 in dimensions, and mapped to CT images by making adjustments on CT images. The corresponding position of the nodules on CT images then can be found on PET images immediately.. 25 .

(32) (a). (b). (c). Figure 3- 2 The axial, coronal and sagittal cross sections of a pulmonary nodule from top to bottom. (a) CT image; (b) PET image; (c) merged PET/CT image. The highlighted regions as shown in the figure are the pulmonary nodules with higher SUV. Computer-aided diagnosis technology such as X-ray and MRI assists physicians to interpret medical images in radiology. Physicians diagnose the patients by viewing the axial, coronal and sagittal cross sections of a pulmonary nodule. The merged image of PET/CT was shown in Figure 3-2. The highlighted regions as shown in the figure are the pulmonary nodules with higher SUV.. 3.4 Image Segmentation Image segmentation is an easy process to analyze a digital image. It partitions digital image into sets of pixels, then we can undergo computation on each partition of the image. Clinical physicians usually compute the maximum value of SUV (SUV max) of the lesion area for pathology decision. An 26 .

(33) alternative way is to calculate the average SUV of region of interest (ROI). However, the position, size and shape of individual tumors are different, therefore it is important to select a right region in the experiment. We obtained the DICOM file from NCKU College of Medicine in Tainan City, Taiwan. We then analyzed the images of CT and PET and merged them into PET/CT images, and finally selected ROI with 31 × 31 pixels. The 31 × 31 pixels of tumors were limited to less than 3 cm in size for 3 cm is the range and limit in our discussion. After reading the CT and PET images from the program, we chose the tumors by adjusting the coordinates of their axial, coronal and sagittal planes, and kept every segmented image for later processing.. 3.5 Preprocessing The imaging data taken from the PET/CT images cannot be directly assessed in the experiment, because it contains a lot of unwanted noise which will interfere with the correct information. Therefore, image preprocessings need to be done in order to reduce background noise and enhance data images prior to computation of variables. We created a threshold value for the boundaries of the nodules as it could be used to leave out unnecessary areas of the nodules. The computer then automatically drew the contour by tracing the coordinate points formed by the images.. 3.5.1 Thresholding . Thresholding is intuitively applied in image segmentation because it is 27 . .

(34) simple in implementation. It accelerates the computation by turning the gray-scale images of pulmonary nodules into binary images which contain only black and white for each pixel. We selected a threshold value of 0.5 to exclude unnecessary areas of the nodules to prevent poor performance during computation.. (a). (b). Figure 3- 3 Thresholded image of a pulmonary nodule. (a) the original image; (b) the thresholded image using a gray-value threshold of 0.5 on image (a). . 3.5.2 Contour In this study, we used a new variational formula as shown in Li, C.’s study for geometric active contours. It consisted of an internal energy term that penalized the deviation of the level set function from a signed distance function for object boundaries by minimizing the overall functional energy. The variational level set formulation has the following advantages: speeding up the curve evolution, the general functions to initialize the level set function are more easily to construct than the signed distance function, and can be easily implemented (Li, C., et al., 2005). As shown in Figure 3-4, indented coordinate points were set up on the 28 .

(35) image after thresholding. The indented coordinate points were joined together to form a circular red line which surrounds the tumor boundary. The red line was iterated along the tumor boundary according to its internal energy term, until a clear dividing line was caught to obtain the contoured image.. (a). (b). (c). (d). (e). (f). (g). Figure 3- 4 The image contouring of a pulmonary nodule. (a) The original image; (b)-(e) a red line was iterated to form a circle around the tumor according to its internal energy term, until (f) a dividing line was caught around the tumor boundary; (g) a contoured image of a pulmonary nodule.. Contouring could not be performed well when the tumor boundary is too 29 .

(36) close to the image frame. In such circumstance, a manual dividing line is set up to surround the desired part of the tumor to obtained the contoured image as shown in Figure 3-5.. (a). (b). (c). (d). (e). (f). Figure 3- 5 Manual dividing line set up to obtain a contoured image. (a) the original image; (b)-(c) a red line was iterated to circle the tumor according to its internal energy term; (d) the dividing line exceeded the image boundary; (e) a manual dividing line was set up to surround the desired part of the tumor; (f) a contoured image of a pulmonary nodule.. 3.6 Variable Computation The variables such as mean SUV, max SUV of PET, areas of CT, mean of CT were computed by the formulas stated in Chapter 2 after noise was removed during pre-processing. Physicians in NCKU College of Medicine gave us two different kinds of 30 .

(37) patients’ diagnosis and medical images: benign nodules and malignant nodules. The state of illness and the description of patients’ suspicious nodules were written on the diagnostic report, so we could find the nodules according to diagnostic medical keywords such as “hyper metabolic cavitated nodule”, “lung mass”, “pulmonary nodule”, “lung nodule”, and “lung lesion” as well as the size, position and SUV.. a hypermetabolic cavitated nodule in right upper lung size: 2.2 cm, SUV: 4.8 Figure 3- 6 Images of a segmented malignant pulmonary nodule with its description. Starting from left to right is the original CT image, thresholded and contoured image, and the PET image. Figure 3-6 shows the description of one of the segmented malignant nodules. The full description of segmented malignant nodules is listed in Appendix A. The images and description of the segmented benign nodules in Appendix B are arranged in the same way as shown in Appendix A. The left column of the table consists of three images, starting from the left to right: the original CT image, image after preprocessing and the corresponding PET image of CT image. The right column of the table lists the information obtained from the images in left column. We then used this information to locate the tumor, segment and compute the variables from the related formulas. . 3.7 Variable Selection 31 .

(38) The variables computed from the experiment were applied with GA for classification of tumors. When the variables were applied with GA at generation 0, their frequency of bit 1 is randomly distributed. At 10th generations, the frequencies of some variables become 0, while some become 1 (Figure 3.8). This is known as the local optimum resulting from GA development. As shown in Figure 3.7, the horizontal axis shows the generations of the variables applied with GA, starting from 0 to 10; the vertical axis shows the fitness of the variables, which is the size of the area under ROC, ranging between 0 and 1.. Figure 3- 7 Maximum and average fitness.. 32 .

(39) . (a). (b). Figure 3- 8 The frequency of bit 1 at (a) generation = 0, and (b) generation = 10 resulting from GA development.. 3.8 Cross-validation Cross-validation, also known as rotation estimation, is a practical method of partitioning data sample into smaller subsets in statistics.. A subset is first. analyzed in this process, while others are used for subsequent confirmation and verification based on this analysis. The subset used for analysis is called a training set, while others are called validation set or testing set. Usually multiple rounds of cross-validation are performed over the subsets to receive the average validation results. Common types of cross-validation include K-fold cross-validation, 2-fold cross-validation, repeated random sub-sampling validation and leave-one-out cross-validation. In K-fold cross-validation, the initial sample is partitioned into k subsamples in a random way. A single subsample is retained as testing data, and the other k-1 subsamples are used as training data. Repeating the 33 .

(40) cross-validation k times, once for each subsample validation, a single estimate was obtained ultimately from the average results obtained after k times (Moreno-Torres, J.G. et al., 2010). A 10-fold cross-validation is usually used to receive the average results; we performed a 5-fold cross validation instead due to the small numbers of variables used in our GA and SVM classification.. 34 .

(41) Chapter 4. Experiment Results. Texture calculations were done on the suspicious nodules taken from the PET/CT images, and these calculations were then applied with the genetic algorithm program for feature selection before proceeding with SVM classifier. A set of input variables were randomly selected and evaluated under the GA program at first, then were proceeded to detect the optimal hyper-plane by using SVM classifier for classification of the nodules. We ran 5-fold cross-validation five times for each set of variables processed from GA screening and SVM classification. Each of the 26 variables inputted was performed with 5-fold cross validation. Finally, statistical results of sensitivity, specificity and accuracy were obtained from each experiment and were used to calculate their means and standard deviations.. Sensitivity =. number of true positives / (number of true positives + number of false negatives). =. probability of a positive test, given that the patient is ill. Specificity =. number of true negatives / (number of true negatives + number of false positives). =. probability of a negatives test, given that the patient is well 35 . .

(42) Accuracy rate =. (numbers of benign nodules × sensitivity + numbers of malignant nodules ×. specificity) / total numbers of nodules. Each of the 22 variables from GLCM (not including the first 4 variables) has four distance vectors. These variables can have four different changing distance vectors. If each variable in each distance vector are put together, the number of variables will be even more than the number of items in experimental nodules, and it cannot properly find the most appropriate variables for screening. Therefore, experiments were done by separating these variables in different directions. We repeated experiments, five times in each direction, found out the sensitivity, specificity and accuracy, and calculated the mean and standard deviation, as shown in Table 4-1.. Table 4- 1 Sensitivity, specificity and accuracy, and their mean and standard deviation of classifications in each direction to GA Distance vector [0,1] (1) [0,1] (2) [0,1] (3) [0,1] (4) [0,1] (5) 79.31 68.97 68.97 68.97 72.41 Sensitivity (%) Mean 71.73 Standard Deviation 4.49 77.78 69.44 72.22 69.44 75.00 Specificity (%) Mean 72.78 Standard Deviation 3.62 78.63 69.18 70.42 69.18 73.57 Accuracy (%) Mean 72.20 36 .

(43) Standard Deviation. 4.02 . . Sensitivity (%) Mean Standard Deviation Specificity (%) Mean Standard Deviation Accuracy (%) Mean Standard Deviation. [1,1] (1) 72.41 . 69.44 . 71.09 . Distance vector [1,1] (2) [1,1] (3) 72.41 68.97 68.97 3.45 72.22 69.44 68.89 4.56 72.33 69.18 68.93 3.37. [1,1] (4) 65.52 . [1,1] (5) 65.52 . 61.11 . 72.22 . 63.55 . 68.51 . . Sensitivity (%) Mean Standard Deviation Specificity (%) Mean Standard Deviation Accuracy (%) Mean Standard Deviation. [-1,1] (1) 68.97 . 69.44 . 69.18 . Distance vector [-1,1] (2) [-1,1] (3) [-1,1] (4) 55.17 68.97 72.41 64.83 7.48 55.56 69.44 72.22 65.00 7.50 55.34 69.18 72.33 64.90 7.49. [-1,1] (5) 58.62 . 58.33 . 58.49 . . Sensitivity (%) Mean Standard Deviation Specificity (%) Mean Standard Deviation Accuracy (%) Mean. [-1,-1] (1) 65.52 . 63.89 . 64.79 . Distance vector [-1,-1] (2) [-1,-1] (3) [-1,-1] (4) 62.07 65.52 68.97 66.90 3.93 61.11 69.44 66.67 66.67 4.39 61.64 67.27 67.94 66.79 37 . . [-1,-1] (5) 72.41 . 72.22 . 72.33 .

(44) Standard Deviation. 3.96. . The final statistical results obtained from GA screening were calculated. Each result showed that out of the 26 variables, some were used and some were not used. The results were summed according to their variables. Running five times for different directions, as represented by [0, 1], [1, 1], [-1, 0] and [-1, -1] respectively, and each time there will be cross validation five times in each SVM. Therefore, each variable appear at most 25 times in each direction as shown in Table 4-2 which shows the number of occurrences of each variable.. Table 4- 2 The number of occurrences of each variable number 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19. Variable name. [0,1] 6 25 7 5 22 19 11 3 13 10 7 4 10 7 4 7 13 12 5. Mean SUV of PET Max SUV of PET Areas CT Mean Contrast Correlation(m) Energy Homogeneity(m) Correlation, Homogeneity Entropy Variance Sum entropy Cluster Prominence Cluster Shade Autocorrelation Dissimilarity Maximum probability Sum average 38 . . Occurrence [1,1] [-1,0] 4 4 23 23 7 11 8 7 16 9 6 5 11 11 11 14 12 7 7 17 7 7 7 5 4 8 5 1 1 5 4 7 10 17 15 13 6 5. [-1,-1] 8 25 4 10 11 3 12 6 9 14 9 5 5 5 5 9 16 11 3.

(45) 20 21 22 23. Sum variance Difference variance Difference entropy Information measure of correlation1 Information measure of correlation2 Inverse difference normalized (INN) Inverse difference moment normalized. 24 25 26. 5 21 8 16. 8 20 7 17. 4 21 8 17. 6 22 7 13. 18. 15. 16. 17. 6. 11. 8. 5. 25. 21. 19. 23. After calculating the number of occurrences of each variable, we took out variables with occurrences higher than 15, 20 and 22. The total numbers of variables occurring more than 15 times were found to be 23. Those occurring more than 20 and 22 times were found to be 14 and 5, respectively. Table 4-3 listed these variables.. Table 4- 3 The numbers of those higher frequency variances Number occurrences. Variables. of variables. Max SUV of PET, [0,1] Contrast, Correlation(m), Difference variance, ≥ 15. 23 Information measure of correlation1, Information measure of correlation2, Inverse difference moment 39 . .

(46) normalized [1,1] Contrast, Maximum probability, Difference variance, Information measure of correlation1, Information measure of correlation2, Inverse difference moment normalized [-1,0] Homogeneity, Dissimilarity, Difference variance, Information measure of correlation1, Information measure of correlation2, Inverse difference moment normalized [-1,-1] Dissimilarity, Difference variance, Information measure of correlation2, Inverse difference moment normalized Max SUV of PET [0,1] Contrast, Difference variance, Inverse difference moment normalized [1,1] Difference variance, Information measure of correlation1, Information measure of correlation2, ≥ 20. 14 Inverse difference moment normalized [-1,0] Homogeneity, Dissimilarity, Difference variance, Inverse difference moment normalized [-1,-1] Difference variance, Inverse difference moment normalized. ≥ 22. Max SUV of PET. 5 40 . .

(47) [0,1] Contrast, Inverse difference moment normalized [-1,-1] Difference variance, Inverse difference moment normalized. These three set of variables of different numbers were applied to classify the original 65 nodules. Because no variable selection was needed, we ran the SVM directly. It was the same as having five times cross-validation and the results were as shown in the Table 4-4. When we applied to SVM with the selected 22 variables, their sensitivity was found to be 72.41%, specificity 72.22% and accuracy 72.30%. If 14 screened variables were used, the sensitivity was 79.31%, the specificity was72.22%, and the accuracy was 75.38%. When 5 screened variables, the sensitivity, specificity and accuracy were 79.31%, 80.56%, and 80.00% respectively.. Table 4- 4 Sensitivity, Specificity and Accuracy of three sets of selected variables using SVM only Occurrences. ≥15. ≥20. ≥22. Numbers of variables Sensitivity (%) Specificity (%) Accuracy (%). 23. 14. 5. 72.41 72.22 72.30. 79.31 72.22 75.38. 79.31 80.56 80.00. Those which appeared more than 22 times showed a greater accuracy rate than which appeared more than 15 and 20 times. On the other hand, the 41 .

(48) numbers of screened variables were less than the original 26 variables, so it was shown that the differentiating efficiency did not depend on the number of variables, but on its combination of variables. Special combination could improve the accuracy of identifying benign and malignant nodules.. 42 .

(49) Chapter 5. Conclusion. We had presented that by using a special combination of variables, the accuracy in distinguishing between benign and malignant tumors was increased. Variables such as mean SUV and SUVmax of PET, area of CT and its mean, and the texture characteristics from GLCM were applied with GA and SVM classifier to produce the most appropriate factors for differentiation. Our current study implemented a total of 26 variables to apply with GA and SVM. The co-occurrence matrix had four distance vectors which would generate four times the numbers of variables for operation. In order to prevent over-training due to the small number of patients, we needed to operate the variables in four distance vectors separately. The number of variables which is much than the total number of image would decrease the accuracy rate. The future direction is that researchers can investigate whether the adding of more screening variables can produce a better combination of variables. Besides texture feature, other features such as shape features, statistical features and characteristics of the relevant position can also be taken into account. These will make computer aided diagnosis (CAD) easily to distinguish the nodules which could not be diagnosed with unaided eyes. For further implementation, the combination of variables can be applied in medical imaging diagnosis. Physicians can manually select any suspicious nodules from PET/CT images and let the computer compute the screened variables immediately to determine whether the particular solitary nodule is benign or malignant. The accuracy of diagnosis can thus be improved and it will 43 .

(50) bring physicians benefits in PET/CT imaging cancer diagnosis.. 44 .

(51) References 方惠珊，(2007)，肺腺癌細胞之電腦輔助圖形辨識，碩士論文，國立雲林科技大學資訊管理系行政院衛生署 Department of Health, Executive Yuan, R.O.C. (TAIWAN)，101 年死因統計結果分析 (2013/06/06)，http://www.doh.gov.tw/CHT2006/DisplayStatisticFile.aspx ?d=88577&s=1 陳玥汝，(2012)，使用透視法於肺部PET/CT DICOM 影像之三維立體顯示，碩士論文，國立高雄大學資訊工程所碩士班曾林維，(2006)，Texture analysis of computed tomography images in acute middle cerebral artery ischemic stroke (利用紋理分析於電腦斷層影像偵測急性缺血性中大腦動脈中風)，碩士學位論文，中原大學醫學工程學系黃俊忠，(2006)，The mass detection in mammography using feature images (利用X光乳房攝影產生之紋理特徵影像在腫瘤偵測上之研究)，碩士論文，國立中央大學資訊工程研究所楊勝智，(2006)，Computer-aided system for diagnosis of masses in breast medical images (乳房醫學影像之腫瘤電腦輔助診斷系統)，博士論文，國立成功大學電機工程學系鄭煥勳， (2006) ， Feature analysis for lung nodule in dynamic contrast enhancement CT image (肺臟腫瘤於動態顯影CT影像之特徵分析)，碩士論文，中原大學醫學工程學系賴鄭婷，(2008)，Integrating the PSONN and Boltzmann function for feature 45 .

(52) selection and classification of Lymph node in ultrasound images(結合粒子群體最佳化神經網路與波茲曼函數應用於超音波淋巴結影像之最佳化特徵選取及分類)，碩士論文，國立雲林科技大學資訊工程研究所碩士班 Albregtsen, F., (2008), Statistical texture measures computed from gray level coocurrence matrices, Image Processing Laboratory Department of Informatics University of Oslo. Clausi, D.A., (2002), An analysis of co-occurrence texture statistics as a function of grey level quantization, Can. J. Remote Sensing, 28, 45-62. da Silva Sousaa, J.R.F., Silvaa, A.C., de Paivaa, A.C., Nunesb, R.A., (2010), Methodology for automatic detection of lung nodules in computerized tomography images. Computer Methods and Programs in Biomedicine, 98, 1-14. Hall, M.N., Jagannathan, J.P., Ramaiya, N.H., Shinagare, A.B. and Van den Abbeele, A.D., (2010), Imaging of Extraosseous Myeloma: CT, PET/CT, and MRI Features, American Journal of Roentgenology, 195, 1057-1065. Han, D., Bayouth, J., Song, Q., Taurani, A., Sonka, M., Buatti, J. and Wu, X., (2011), Globally optimal tumor segmentation in PET-CT images: a graph-based co-segmentation method. Information Processing in Medical Imaging, 22, 245-256. Haralick, R.M., (1979), Statistical and structural approaches to texture, Proceedings of the IEEE, 67, 786-804. Haralick, R.M., Shanmugam, K. and Dinstein, I., (1973) Textural features for 46 .

(53) image classification, IEEE Transactions on Systems, Man and Cybernetics, SMC-3, 610-621. Hu, S. and Reinhardt, J.M., (2001), Automatic lung segmentation for accurate quantitation of volumetric X-Ray CT images. IEEE Transactions on Medical Imaging, 20, 490-498. Khalaf, M., Abdel-Nabi H., Baker, J., Shao, Y., Lamonica D. and Gona J., (2008), Relation between nodule size and 18F-FDG-PET SUV for malignant and benign pulmonary nodules, Journal of Hematology & Oncology. Li, C., Xu, C., Gui, C. and Fox, M.D., (2005), Level set evolution without re-initialization: a new variational formulation, IEEE Computer Society Conference on Computer Vision and Pattern Recognition. Marom, E.M., Erasmus, J.J. and Patz Jr, E.F., (2000), Lung cancer and positron emission tomography with fluorodeoxyglucose, Lung Cancer, 28, 187–202. Masa-ah, P., Tuntawiroon, M., Soongsathitanon, S., (2010), A novel. scheme. for Standardized Uptake Value (SUV) calculation in PET scans. International Journal of Mathematical Models and Methods in Applied Sciences, 4, 291-299. McNitt-Gray, M.F., Hart, E.M., Wyckoff, N., Sayre, J.W., Goldin, J.G., Aberle, D.R., (1999), A pattern classification approach to characterizing solitary pulmonary nodules imaged on high resolution CT: preliminary results, The International Journal of Medical Physics Research and Practice, 26, 47 .

(54) 880-888 McNitt-Gray, M.F., Wyckoff, N., Sayre, J.W., Goldin, J.G., Aberle, D.R., (1999), The effects of co-occurrence matrix based texture parameters on the classification of solitary pulmonary nodules imaged on computed tomography, Computerized Medical Imaging and Graphics, 23, 339-348. Mitchell, M., (1999), An Introduction to Genetic Algorithm, Cambridge, Massachusetts: Massachusetts Institute of Technology. Moreno-Torres, J.G., Sáez, J.A. and Francisco Herrera, (2012), Study on the impact of partition-induced dataset shift on k-fold cross-validation, IEEE Transactions on Neural Networks and Learning Systems, 23, 1304-1312. Potesil, V., Huang, X. and Zhou, X.S., (2007), Automated tumour delineation using joint PET/CT information, Medical Imaging 2007: Computer-Aided Diagnosis, 6514. Rangayyan, R.M., Guliato, D., de Carvalho,J.D, and Santiago, S.A., (2006) Feature extraction from the turning angle function for the classification of contours of breast tumors. IEEE Special Topic Symposium on Information Technology in Biomedicine, Iaonnina, Greece, 4 pages CDROM. Soh, L.K. and Tsatsoulis, C., (1999), Texture analysis of sar sea ice imagery using gray level co-occurrence matrices, IEEE Transactions on Geoscience and Remote Sensing, 37, 780-795. Zsoter, N., Bandi, P., Szabo, G., Toth, Z., Bundschuh, R.A., Dinges, J. and Papp, L., (2012), PET-CT based automated lung nodule detection, 34th Annual International Conference of the IEEE EMBS. 48 .

(55) 49 .

(56) Appendix A The table below shows the description of malignant nodules, starting from the left to right: the original CT image, image after preprocessing and the corresponding PET image of CT image. a hypermetabolic cavitated nodule in right upper lung size: 2.2 cm, SUV: 4.8 lung mass in the right upper lung size: 3.3 cm, SUV: 6.6 pulmonary nodule in the right upper lung size: 2.0 cm, SUV: 4.8 a lung nodule in the left upper lung size: 2.3 × 1.5 × 1.7 cm, SUV: 6.8 a hypermetabolic nodule in the right upper lung 3.7 × 2.2 × 2.1 cm, SUV: 6.0 a lung nodule in the left upper lung size: 2.2 cm, SUV: 5.3. 50 .

(57) a lung mass in the left upper lung size: 4.3 × 3.5 × 2.8 cm, SUV: 11.5 a hypermetabolic tumor in the right upper lung size: 2.7 × 2.4 × 2.7 cm, SUV: 8.6 small pulmonary nodule in the right middle lung, size: 0.4 cm a pulmonary nodule in the right lower lung size: 1.2 cm, SUV: 4.3 a pulmonary mass in the left lower lung size: 3.5 cm, SUV: 5.9 a pulmonary nodule in the left upper lung size: 2.5 cm, SUV: 4.7 lung nodule in the RUL size: 1.7 cm, SUV: 3.8. a lung mass in the right upper lung size: 3.5 × 2.4 × 2.3 cm, SUV: 5.0. 51 .

(58) a hypermetabolic nodule in the left upper lung size: 2.4 cm, SUV: 2.9 a hypermetabolic mass in the right upper lung 4.0 × 3.1 × 2.7 cm in size, SUV: 2.2 a pulmonary nodule in the right upper lung size: 1.7 cm, SUV: 3.1 a lung nodule in the left upper lung size: 1.5 cm, SUV: 6.1. a lung nodule in the left upper lung size: 1.6 cm, SUV: 3.7 a FDG-avid lung mass in the right upper lung size: 3.4 × 2.3 × 2.4 cm, SUV: 11.7 a small pulmonary nodule in the right lower lung, adjacent pulmonary tissue (0.8cm) a pulmonary lesion in the right upper lung size: 1.2 cm, SUV: 1.4 52 .

(59) a pulmonary nodule in the RLL size: 1.3 cm, SUV: 4.1 two pulmonary nodules one in the LLL size: 0.9 cm, SUV: 1.4 a hypermetabolic nodule in the right upper lung size: 4.3 × 3.3 × 3.3 cm, SUV: 7.6 a lung tumor in the left upper lung size: 3.5 × 2.8 × 2.4 cm, SUV: 4.4 a hypermetabolic nodule in the left upper lung size: 2.0 × 1.6 × 1.9 cm, SUV: 6.2 lung mass in RML size: 4.3 × 3.1 × 2.7 cm, SUV: 10.3. a lung tumor in the right upper lung size: 3.5 × 2.7 × 2.5 cm, SUV: 10.4. 53 .

(60) Appendix B The table below shows the description of benign nodules, starting from the left to right: the original CT image, image after preprocessing and the corresponding PET image of CT image. sella mass, with volume 2.64 ml, length 2.0 cm, SUV (initial) 8.28 a hypermetabolic nodule in the left lower lung, 2.9 × 2.1 × 1.7 cm in size, SUVmax: 3.8 a nodule, 1.0 cm, SUV 1.2, in left lower lung. a lung mass in the right lower lung size: 4.7 × 3.0 × 2.5 cm, SUV: 4.3. in LUL mass [4.8cm], with volume 6.44 ml, SUV 0.88. a patchy lesion in the right upper lung size: 1.1 cm, SUV: 0.8. 54 .

(61) ground glass lesion in the right middle lung. left upper lung is absent due to previous operation a hypermetabolic mass in the right upper lung size: 4.1cm, SUV; 9.0 a hypermetabolic nodule in the right upper lung, 2.4 cm in size, SUV: 4.2. in the 1.0 cm LLL lung mass,with volume 0.39 ml, SUV 0.44 seen on CT images, with indistinguishable FDG uptake from adjacent pulmonary tissue one is in the RML (size: 0.7 cm) and another is in the RUL (size: 0.4 cm). A lobulated nodule in right lower lung size: 3 cm, SUV: 2.4. 55 .

(62) Multiple enlarged lymph nodes, up to 1 cm SUV up to 4.8, in right lower interlobar, both hilar para-tracheal a lung nodule in the right upper lung size: 1.5 cm, SUV: 3.6 a pulmonary nodule in the left upper lung size: 1.5 cm, SUV: 6.2 a tumor in left upper lung size: 1.8 cm, SUV 6.8. a furhter increase of FDG, SUV 8.2, is noted on delayed scan.. a patch of mild FDG uptake, SUV 1.4, in right middle lung.. in RUL nodule, with volume 1.16 ml, length 1.3 cm, SUV 1.75. in left upper lung lesion size: 2 cm, SUV 2.3,. 56 .

(63) A patchy infiltration with mild FDG uptake, SUV 2.2, in superior segment of left lower lung. pulmonary lesions in the right lower lung size: 2.0 cm, SUV: 1.1 pulmonary lesions in left lower lung size: 0.8 cm, SUV: 1.3. a lung mass in the right lower lung size : 6.2 × 4.1 × 3.2 cm, SUV: 3.7. nodule in right lower lung size : 1.8 cm, SUV 2.5. a small RLL lung nodule size : 0.9 cm, SUV: 1.7. relatively decreased cerebral cortical uptake of FDG is observed. a pulmonary nodule in the left upper lung size: 1.1 cm, SUV: 1.0 57 .

(64) a lung nodule in the right lower lung size: 1.5 cm, SUV: 1.0 pulmonary nodule in the right upper lung size: 1.5 cm, SUV: 2.6 a lung mass in the right middle lung size: 3.6 × 2.6 × 3.1 cm, SUV: 4.9.. a nodule in left lower lung size: 2.2 cm, SUV 4.5 a hypermetabolic nodule in the right main bronchus size: 3 × 4 × 2.4cm, SUV 10.3 in RUL lung nodule, with volume 2.23 ml, SUV: 4.08. 58 .

(65)