動態乳房彈性超音波之腫瘤偵測與分析

(1)

國立臺灣大學電機資訊學院資訊工程學系碩士論文

Department of Computer Science and Information Engineering College of Electrical Engineering and Computer Science

National Taiwan University Master Thesis

動態乳房彈性超音波之腫瘤偵測與分析 Tumor Analysis of Dynamic Elastography

黃冠穎

Guan-Ying Huang

指導教授：張瑞峰博士 Advisor: Ruey-Feng Chang, Ph.D.

中華民國 98 年 7 月 July, 2007

(2)

i

ACKNOWLEDGEMENTS

First, I have to sincerely appreciate my advisor, Dr. Ruey-Feng Chang, for his patient guidance and discussions. I also thank the physician Dr. Moon(Seoul National University Hospital, Korea) for providing the elastography data and the sufficient information for the data analysis. I also want to thank all fellows in Medical Image Processing lab at the Department of the Computer Science and Information Engineering, National Taiwan University for the for their helps and encouragement during preparing this thesis.

Finally, I thank my family and friends for the encouragement and company. I dedicate this thesis to everyone who has assisted me in these two years.

(3)

ii

動態乳房彈性超音波之腫瘤偵測與分析

研究生：黃冠穎指導教授：張瑞峰博士

國立台灣大學資訊工程研究所

摘要

近年來，乳房彈性超音波已成為測量腫瘤彈性最常見的方法，醫生必須施以輕微的壓力於腫瘤組織之上，藉以得到一連串連續的動態彈力影像。而醫生將會從這一序列的影像中挑選出最具代表性的一張影像，此張影像之品質優劣亦將決定診斷的準確性。此篇論文之目的為量化彈力超音波影像的品質並選出彈力超音波影片中最具代表性的一張影像，同時也將實作一可半自動切割腫瘤以利計算腫瘤的彈性比率。利用使用者在首張影像所選取的種子點，配合邊緣偵測(Edge Detection)以及區域增長(Region Growing)的方法可自動地切出整個資料中的腫瘤輪廓，此外，種子點會根據之前影像中腫瘤移動的方式做相同的位移，如此方可得到較好的切割結果。根據所得的腫瘤輪廓中的一致性以及腫瘤跟正常組織的對

比性，可以量化兩種彈力影像品質：信號雜訊比（SNRe）以及對比雜訊比（CNRe），

並根據量化結果挑選出最具代表性之影像作為判斷良惡性之用。本實驗中以141

個經過病理驗證的病例進行測試，包含93 個良性以及 48 個惡性的病例，比較使

用本篇論文的方法所選的影像以及所得彈力最差的影像、壓力最大時的影像和醫

生所選之影像，並計算Mann-Whitney U 測試、效能以及 ROC 曲線來評估結果。

經由實驗，SNRe的準確率為 84.40%，敏感度為 83.33%，特異性為 84.95%，而

(4)

iii

ROC 曲線的 Az 值則是 0.90；CNRe則是有 82.27%的準確率，79.17%敏感度，

83.87%特異性，Az 值為 0.88，兩者均有不錯的效率，因此我們歸納出以此種方式進行影像品質的量化並挑選出最具代表性影像是可行並且較醫生挑選為客觀

的。此外，為了縮減計算分析的時間，亦提出了一種fast-selection 方式來挑選代

表影像，此方法將只針對第一張影像進行腫瘤切割，經測試仍有一定的準確度並且大幅的減少分析時間。

(5)

iv

Tumor Analysis of Dynamic Breast Elastography

STUDENT: GUAN-YING HUANG ADVISOR: DR. RUEY-FENG CHANG

INSTITUTE OF COMPUTER SCIENCE AND INFORMATION ENGINEERING NATIONAL TAIWAN UNIVERSITY

Abstract

Recently, the sonoelastography has been the most general technique to measure the tumor strain. In the sonoelastography, the physicians need to lightly compress a tumor to obtain a dynamic elastographic image sequence which is composed of continuous elastographic slice. A representative slice of the dynamic elastographic image sequence will be selected by the physician and the quality of this selected slice will affect the diagnosis result. The purpose of this study is to quantify the elastographic images quality and select a representative slice from an elastography movie file. This study also proposes a semi-automatic segmentation to find the tumor contour for calculating the hard ratio of tumor. Utilizing a group of seeds given by the user in the first slice, the automatic segmentation using the edge-detection and region growing methods is applied in the first slice and then the subsequent slices. Moreover, the seeds of the subsequent slices will be moved according to the tumor displacement to improve the segmentation results. After finding the tumor contours, two quality

(6)

v

quantification methods, the signal to noise ratio of (SNRe) and contrast to noise ratio (CNRe) of elastographic slice, are computed according to the uniformity inside the selected region or the contrast of the tumor and the surrounding normal tissue. Finally, find a representative slice based on the quantification and use the selected slice to differentiate the benign and the malignant lesions. In this study, 141 biopsy-proved sonoelastography composed of 93 benign and 48 malignant masses are used to evaluate the performance of the quantification methods. In the experiments, the diagnosis results of the slices selected by two proposed methods are compared with those of the maximum compression slices, maximum strain slices, and the slices selected by physicians. The Mann-Whitney U test, performance indexes, and receiver operation curve (ROC) are applied to examine the effectiveness of the proposed quantification methods. According to the result of experiment, the accuracy, sensitivity, specificity, and the Az value for the SNRe are 84.40%, 83.33%, 84.95%

and 0.90, respectively and for the CNRe are 82.27%, 79.17%, 83.87% and 0.88, respectively. We can conclude that using the quantification methods to select the representative slice of the elastography is practicable and more objective than that selected by the physician. Moreover, to reduce the run time of the quantification analysis in this paper, a smart fast-selection method is also proposed and only the tumor contour of the selected slice is required to be segmented. The fast-selection method can achieve an acceptable performance and greatly reduce the execution time of the analysis.

(7)

vi

LIST OF FIGURES

Fig. 1 An elastography image. The color-coded elastographic image is on the left

represents the strain information. ... 5

Fig. 2 The circular HSV color space model. The left circular plane is the slice of the model with value 127. ... 6

Fig. 3 A E-score=1 lesion. ... 7

Fig. 8 Flow chart of the proposed system... 10

Fig. 9 The images (a) before (b) after anisotropic diffusion. ... 12

Fig. 10 The seed region shown in the dash lines is the bounding box of three seed points which are shown in black ones. ... 13

Fig. 11 The result images of sigmoid filter with different β values. The image (a), (b), (c), (d) is using mean + 0.5×standard deviation, mean + standard deviation, mean + 1.5×standard deviation, mean + 2×standard deviation as β value of sigmoid function, respectively. ... 13

Fig. 12 The image (a) before (b) after the gray scale closing. ... 14

Fig. 13 The kernels of Sobel edge-detection. The sign “+” means the center of the mask. ... 15

Fig. 14 The image (a) before (b) after the edge detection. ... 15

Fig. 15 The region growing result of Fig. 14-(b), the white region is the tumor region. ... 16 Fig. 16 The displacement of two slices i and (i-1). The black points are the centroid

(9)

viii

of the tumor in the two slices and the displacement is shown in dash line. ... 17 Fig. 17 The regions to quantify of elastography quality. The solid 10×10 rectangle is used to compute the SNRe and the all three rectangles are used to compute the CNRe. The x1 is the distance from the image border to the left boundary of tumor and the x2 is the one from the right boundary of tumor to image border.

The distance from image border to the top boundary of tumor is represented by y. The upper-left corners of (x1/2)×y and (x2/2)×y background blocks B1 and B2

are located at (x1/4, y/2) and (M-3×x1/4, y/2), respectively, where M is the image width. ... 18 Fig. 18 Two slices in a dynamic elastographic image. (a) The slice with low SNRe

value = 48.9362 and (b) The slice with high SNRe value = 153. 6835. The white rectangle represents the calculation region of SNRe. Note that the SNRe values are calculated from the hue information of the elastographic slices. ... 19 Fig. 19 Two slices in a dynamic elastographic image. (a) The slice with low CNRe

value = 0 and (b) The slice with CNRe value = 12.2662. The rectangles are the calculation region of CNRe. Note that the CNRe values are calculated from the hue information of the elastographic slices. ... 20 Fig. 20 The relationship between hue value and strain. The threshold of strain is

about 200. ... 20 Fig. 21 An example to calculate the compression displacement. The distance d is

equal to the compression displacement. ... 24 Fig. 22 The population pyramid of the hard ratio distribution in (a) Rmax (b) Rsnr (c) Rcnr (d) Rcomp (e) Rphy. ... 26 Fig. 23 The ROC curves of all hard ratios obtained by five methods. ... 29 Fig. 24 The user-selected points of the fast-selection method. The central point is

(10)

ix

used in SNRe and all the three points are used to define the background regions in CNRe. ... 31 Fig. 25 The population pyramid of the (a) Rfsnr and (b) Rfcnr. ... 32 Fig. 26 The ROC curves of the ratios using fast-selection method. ... 34 Fig. 27 A true negative example. (a)(b) The elastographic image and the B-mode

image with the segmentation result of the same slice. (c) The slice selected by maximum hard ratio. The hard ratio of this slice is 12.24%. (d) The slice selected by maximum SNRe, 28.77. The hard ratio of this slice is 3.36%. (e) The slice selected by maximum CNRe, 13.18. The hard ratio of this slice is 3.44%. (f) The slice selected by the maximum compression and the hard ratio of the slice is 3.44%. (g) The physician-selected image with the hard ratio 3.29%. (h) The slice selected by the fast-selection using SNRe. The hard ratio is 0.58%. (i) The slice selected by the fast-selection using CNRe. The hard ratio is 12.85%. (j) The curves of different quantification methods. Each value of the quantification is normalized with the maximum value. ... 37 Fig. 28 The false positive example. (a)(b) The elastographic image and the B-mode image with the segmentation result of the same slice. (c) The slice selected by maximum hard ratio. The hard ratio of this slice is 84.59%. (d) The slice selected by maximum SNRe, 70.08. The hard ratio of this slice is 69.32%. (e) The slice selected by maximum CNRe, 12.57. The hard ratio of this slice is 61.31%. (f) The slice selected by the maximum compression and the hard ratio of the slice is 34.08 %. (g) The physician-selected image with the hard ratio 65.83%. (h) The slice selected by the fast-selection using SNRe. The hard ratio is 63.66%. (i) The slice selected by the fast-selection using CNRe. The hard ratio is 70.61%. (j) The curves of different quantification methods. Each value

(11)

x

of the quantification is normalized with the maximum value. ... 39 Fig. 29 A true positive example. (a)(b) The elastographic image and the B-mode

image with the segmentation result of the same slice. (c) The slice selected by maximum hard ratio. The hard ratio of this slice is 82.45%. (d) The slice selected by maximum SNRe, 102.59. The hard ratio of this slice is 82.06%. (e) The slice selected by maximum CNRe, 10.60. The hard ratio of this slice is 66.3%. (f) The slice selected by the maximum compression and the hard ratio of the slice is 68.25%. (g) The physician-selected image with the hard ratio 69.75%. (h) The slice selected by the fast-selection using SNRe. The hard ratio is 72.12%. (i) The slice selected by the fast-selection using CNRe. The hard ratio is 63.93%. (j) The curves of different quantification methods. Each value of the quantification is normalized with the maximum value. ... 41 Fig. 30 A false negative example. (a)(b) The elastographic image and the B-mode

image with the segmentation result of the same slice. (c) The slice selected by maximum hard ratio. The hard ratio of this slice is 19.29%. (d) The slice selected by maximum SNRe, 120.78. The hard ratio of this slice is 13.69%. (e) The slice selected by maximum CNRe, 14.88. The hard ratio of this slice is 14.3%. (f) The slice selected by the maximum compression and the hard ratio of the slice is 14.70%. (g) The physician-selected image with the hard ratio 0%. (h) The slice selected by the fast-selection using SNRe. The hard ratio is 5.55%. (i) The slice selected by the fast-selection using CNRe. The hard ratio is 11.49%. (j) The curves of different quantification methods. Each value of the quantification is normalized with the maximum value. ... 43

(12)

xi

LIST OF TABLES

Table 1 The median value and p-value (using Mann-Whitney U test) of each hard ratio. All the ratios are non-normal distribution using the Kolmogorov-Smirnov test. ... 27 Table 2 Performance indexes of the proposed hard ratios. ... 28 Table 3 The median value and p-value (using Mann-Whitney U test) of Rfsnr and Rfcnr. ... 33 Table 4 The performance indexes of Rfsnr and Rfcnr. ... 33 Table 5 The p-values of the ROC curves between the original quantification methods

and the fast-selection methods. ... 34 Table 6 The p-values of the performance indexes between the original quantification methods and the fast-selection methods. ... 35

(13)

1

Chapter 1 Introduction

According to the report of American Cancer Society, the breast cancer is the first rank in women of estimated new cases in 2009 [1]. The report indicates that there are 192,370 new estimated cases and 40,170 estimated deaths in female breast cancers which is the second rank of estimated deaths. The United States Preventive Services Task Force (USPSTF) recommends that women after 40 years of age should take the examination annually. Early detection can extremely improve the survival rate of the breast cancer [2].

The elastography is a newly developed ultrasound (US) based technique. The ultrasound is a commonly used breast cancer diagnosing technique with many advantages such as convenient usage, real-time capability and non-invasive [3].

Except these advantages, the elastography can be used to evaluate the stiffness of soft tissues [4, 5]. The principle of elastography is to estimate the hardness difference between normal tissues and lesion tissues by measuring the tissue strain induced by the probe compression. According to the clinical studies, the elastography could be used to differentiate the malignant breast masses from the benign masses [6-8].

Generally, the tissue strain is higher in the normal tissues than in the lesion tissues.

This difference would lead to the areas of the benign tumors determined on elastographic image are always the same or smaller than the areas of the tumor on B-mode image, whereas the areas of the malignant tumors on elastographic image are always larger than they are shown on B-mode image [4, 7].

(14)

2

There are two different ways, color or gray scales, to demonstrate the elastographic image in current commercial elastography equipments. The equipment used in this paper converts the strain information into the color scale image and superposes it on the B-mode image. The advantage is easy to recognize the strain of the lesion region. Another type of equipments demonstrates the pure strain image by using the dark colors for hard tissue and the bright colors for soft one [9, 10].

Instead of using the strain information directly, Itoh et al. proposed a score system which classifies the elastographic images into five categories according to the distribution of hard tissue. The better cutoff point of the scores to differentiate the benign and the malignant tumor is between 3 and 4. The experiment result shows that the best sensitivity 86.5% and specificity 89.8% occur in this cutoff [11].

Elastography has been the one of the most effective tool to differentiate between the benign and malignant masses. However, the image quality will not be well in all slices of the elastography data. There could be much artificial noise and elasticity noise which will cause the erroneous judgment. In general, the better slice is selected by physician according to his own experiences. However, this way is subjective because the different physician could choose a different slice. In this paper, a quantification method of quality will be investigated and the aim is to propose a more objective method to automatically select a representative slice to obtain a better diagnosis result.

There are already several methods to quantize the image quality but these methods refer to the information from the radio frequency (RF) data, not from the image [12-14]. The methods stated in this paper propose the image quality quantification according to the information from the elastographic image. There are two quantification methods in this paper; one is the elastography signal-to-noise ratio

(15)

3

(SNRe) and the other is the elastography contrast-to-noise ratio (CNRe). Both these two methods focus on the uniformity of images, but the SNRe pays more attention to the tumor region and the CNRe also refers to the background information.

In order to examine whether the slice with the maximum image quality quantification value could be selected as the representative slice for a case, all tumors in the slices in a movie file of this case are segmented and computed the diagnosis feature. In this paper, the strain rate of a tumor is proposed to diagnosis a tumor. This rate is similar to the score system proposed by Itoh et al. [11]. Furthermore, in order to avoid manually segmenting all the slices, a semi-automatic segmentation of the B-mode image in elastography data is also proposed. The segmentation only needs user to select some seeds in the first slice and then the system will segment all slices in the elastographic movie file based on the edge information. In the experimental results, five strain ratios computed from the four slices with maximum SNRe, maximum CNRe, maximum compression, and maximum strain ratio and the physician-selected slice are compared.

The details of this paper are stated as follows. The image data properties and some related works are introduced in Chapter 2. In Chapter 3, the methods of segmentation and quantification will be presented. The experiment results are discussed in Chapter 4. Finally, the conclusion and the future works are stated in Chapter 5.

(16)

4

Chapter 2 Material and Related Works

Before stating the detail of the proposed system, the properties of the image data should be introduced. In this chapter, the ultrasound system which estimates the strain and records the elastography images is described first. Then, a HSV color model would be discussed for extraction the elasticity information from the color elastography images. Finally, a conventional scoring system of elastography is presented.

2.1 Data Acquisition

Elastography is recently developed method which can present the estimated elasticity information [15]. In the elastography system, two original radio frequency (RF) data are recorded before and after the compression. Subtracting the pre-compressed data from the post-compressed data produces an elastogram. The value of each pixel means how far it shifts and represents the strain of the pixel [16, 17].

The EUB-8500 scanner (Hitachi Medical, Tokyo, Japan) with a 14–6 MHz linear transducer is used to obtain the sonoelastographic and conventional US images in this paper. The lesion is compressed by the transducer with light pressure instead of higher pressure which will cause the manifest nonlinear properties of elasticity [11].

For the elastography image in Fig. 1, the color-coded elastographic image is shown in the left and the B-mode image is shown in the right. The elastographic

(17)

5

image is a B-mode image superposedwith the color scale which uses the 256 color mapping to represent estimated strain. The degree of strain is from red (highest strain), green (intermediate strain), to blue (no strain) [9].

Fig. 1 An elastography image. The color-coded elastographic image is on the left represents the strain information.

2.2 HSV Color Space

In the Hitachi color elastography, the strain value is translated to the color information superposed over the gray scale B-mode image by saving the elasticity information in the hue channel of HSV color space [18]. The HSV stands hue, saturation and value. Comparing to the RGB (Red-Green-Blue) color space, the HSV is a color space closer to the human vision. The method for translating RGB to HSV is shown as follows

H cos

^{R G} ^{R B}

R G R B G B (2.1)

(18)

6

S 1

R G B

min R, G, B

(2.2)

V R G B

(2.3)

A HSV model shown in Fig. 2 is based on the circular color planes which are perpendicular to the Value axis. The Hue refers to the wavelength of color in the form of angle from 0 to 360. The Saturation is the purity of the color and proportional to the distance between the color and the Value axis. In the other word, the gray scale colors like white or black are with S=0 and the pure color like red, green and blue are with S=1. The intensity information is represented by value that the black is 0 and white is 255.

Fig. 2 The circular HSV color space model. The left circular plane is the slice of the model with value 127.

2.3 Elastography Scoring System

Generally, the breast imaging reporting and data system (BI-RADS) is used to evaluate the tumor in B-mode images [19]. In elastography, there is also a similar scoring system based on the strain distribution inside the tumor proposed by Itoh et al.

S H V

Red Green

Blue

Black White

(19)

7

[11, 20]. The E-score scoring system is defined as

Score 1: All tissues inside the hypoechoic lesion are high elasticity (i.e., all pixels in tumor region are green) (Fig. 3).

Score 2: Most of the lesion tissues are elasticity and some areas are solid (e.g.

fibroadenoma) (Fig. 4).

Score 3: Only the periphery region is elasticity and the center of lesion is no strain (Fig. 5).

Score 4: No strain in the entire lesion (i.e., all pixels inside the tumor are shaded in blue) (Fig. 6).

Score 5: The entire lesion and surrounding region are no strain (Fig. 7).

There is a point needed to notice that a lesion with the score of 4 or 5 is more probable to be classified to a cancer. However, the lesion with lower score still has probability to be a cancer. However, to choose the one slice with best elasticity quality could reduce the probability of erroneous judgment.

Fig. 3 A E-score=1 lesion.

(20)

8

(21)

9

Chapter 3 Dynamic Elastography Tumor Segmentation and Analysis

In order to analysis the strain rate of tumor, the tumor contour should be found at first. For a dynamic elastographic movie file, each slice should be extracted and then the tumor contour is required to be found for each slice. An automatic segmentation will be proposed for all the slices except the first one in order to reduce the physician’s efforts. At first, the physician only needs to select some seeds inside the tumor in the first slice and then the tumor contour will be segmented by the computer. For the subsequent slices, the segmentation seeds will be automatically calculated from the seeds in the previous slice. The flow chart of the proposed system is illustrated in Fig. 8. First, the proposed system will extract each elastographic slice from the elastography movie file which was saved in the AVI format and then the strain information is recovered from the color-coded elastographic image. Second, some image processing techniques, like the anisotropic filter [21, 22], sigmoid [23, 24], grayscale morphological closing [18, 25-27] will be applied for improving the segmentation results. The major parts of segmentation are composed of the edge detection [18, 28-31] and region growing [28, 32-35]. Another point should be noticed is that the seed positions have to be modified between two neighboring slices for achieving the better segmentation results. After the segmentation, the system finds the different strain ratios of frames with maximum SNRe [12-14, 36, 37], maximum CNRe [12-14, 37], maximum compression, maximum strain ratio, and the frame chosen by the physician. These five ratios can be used to differentiate benign and

(22)

10

malignant and their accuracy rates will be compared in the experiments described in next chapter.

Fig. 8 Flow chart of the proposed system.

3.1 Dynamic Elastographic Image Segmentation

In order to obtain a better strain ratio, a more approximate tumor region is needed. An appropriate pre-processing can improve the image quality so that the segmentation could be easier and the result of segmentation would be better. With a good pre-processing result, the segmentation method used in this proposed system could be simpler and faster. In this paper, an edge-detection method is applied to get the boundary information and then a region growing is used to obtain the complete region of tumor.

Dynamic elastographic

Segmentation

Computing hard Pre-processing

Computing SNRe Computing CNRe

Seed prediction

(23)

11

3.1.1 Pre-processing

In general, the noise will affect the edge information and make the system obtain a wrong region. Sometimes, the intensity of tumor would be very close to the intensity of shadows behind the tumor. Therefore, the pre-processing is needed to reduce the noise and enhance the contrast between the tumor and background. Three image processing methods, anisotropic diffusion, sigmoid filter [23, 24] and grayscale morphological closing [18, 25-27] are performed on B-mode image.

A. Anisotropic Diffusion

To reduce the noise, some smooth filters like Gaussian smooth [18, 38, 39] are used generally. However, the boundary information would be blurred in the same time;

however, the edge information is important in the segmentation of proposed system.

The anisotropic filter [22] is an edge preserving smooth filter and it can satisfy the requirement of reducing the noise but preserving the edges. The diffusion function of anisotropic diffusion is shown as (3.1).

div , , , , ∆ (3.1)

where I(x,y,0) is the input image and I(.,t) are a family of images. The symbol and

∆ means the gradient and Laplacian operations, respectively. The function c is the conductance parameter as

| | ^{| |} (3.2)

where k is a coefficient to control the sensitivity of edge. The function will reduce the conductance while I, which represents the gradient magnitude, is large. The result of anisotropic filter is illustrated in Fig. 9.

(24)

B. S

high a pr cont

whe para ima whi rang Fig.

auto dev dev plus cho

F

Sigmoid Enh Although her, there ar roper boun trast of ima

ere I’ is the ameters Ma age, and are ich is relate ge. A seed r . 10 and t omatically d

iation of th iation. The s one stand

ice of β valu

(a Fig. 9 The im

hancement the intensi re still many ndary of tum age. The fun

′

intensity o ax and Min e set to 255 ed to the w region will b

the mean a decide α an he seed reg results for dard deviati

ue is mean

a) mages (a) be

ity of tumo y cases who mor. The s nction is for

f result ima represent th 5and 0 gene width of inte be selected and standar nd β values.

gion and the different β on has bett plus on stan

12

efore (b) aft

or is lower ose contrast sigmoid fun rmulated as:

·

age and I is he maximum erally. The m ensity range

according t rd deviatio . In this pap e β value i β values are ter result to ndard devia

( fter anisotrop

and the int t is so low t nction could

:

the intensit m and mini most impor e and β is o the seeds’

n of this per, the α v is set to the e shown in o isolate the ation.

(b)

pic diffusio

tensity of n that the syst d be used

ty of input i imum value rtant two pa the center

’ locations l seed region value is set

e mean plu Fig. 11 and e tumor. He on.

normal tissu tem cannot to enhance

( image. The es of the ou

arameters a of the inten like

n will used to the stan us one stan d the β = m ence, the b

ue is find e the

3.3) two utput are α nsity

d to ndard ndard mean etter

(25)

Fig.

C. G

back too

. 10 The se points

. 11 The re (c), (d mean sigmo

Grayscale M Even if kground, th close to th

eed region s which are

(a

(c sult images d) is using m

+ 1.5×stan oid function,

Morphologic the sigmoi here are som he intensity

shown in t shown in bl

a)

c) of sigmoid mean + 0.5 dard deviat , respective

cal Closing id filter ca me situation

values of t

13

the dash lin lack ones.

d filter with

×standard d tion, mean

ly.

an enhance ns that the i tumor. The

nes is the bo

(b

different β deviation, m

+ 2×standa

e the cont intensity va basic idea

ounding bo

b)

(d)

values. The mean + stan ard deviatio

trast betwe alues of sha to solve thi

ox of three

e image (a), ndard deviat n as β valu

een tumor adows or fat is problem

seed

, (b), tion, ue of

and t are is if

(26)

som the [18, dila wou be r tum

3.1.

segm edg regi by e is a used resu

A.

me narrow sh tumor is iso , 25-27], can ation and th

uld do the o removed if mor would be

Fi

2 Segmenta After the mentation

e-detection ion growing edges. On th little differ d. Hence, a ult.

Sobel Edge

hadows con olated. A m

n achieve th hen erosion opposite thin

their width e isolated. T

(a) ig. 12 The im

ation e pre-proces

[28, 33, 3 process can g segmentat

he other han rent in each a seed pred

e Detection

nnect to the morphologica

his purpose.

n. Dilation w ng. Hence, t

are small t The result is

mage (a) be

ssing, an edg 35] are pe n provide th tion can me nd, for the d h slice and diction proc

14

tumor, the al image pro . The graysc would emp the connect than the rad s shown as F

efore (b) aft

ge-detection erformed t he informat erge the par dynamic ela the seeds o cess is nee

connections ocessing tec cale closing phasize the tions betwee dius of the f

Fig. 12.

ter the gray

n [18, 29] a to extract tion about t

rts of tumo astographic of previous eded to ens

s could be b chnique, gra g is compose white secti en shadows filter kernel

(b) scale closin

and a region the tumor tumor conto r if the tum

images, the slice could sure a bette

broken and ay scale clo ed of gray s ion but ero s and tumor . Therefore

ng.

n growing b r contour.

our and then mor is separ

e tumor con d not be dire er segmenta

then osing scale osion will , the

ased The n the rated ntour ectly ation

(27)

time Sob of S repr Com

The 14.

ima com

Fig.

The edge e and there bel edge det

Sobel opera resents the mbining the

en, binarize The tumor age. Therefo mplete conto

. 13 The ke mask.

e detection are many d ection whic ators is desc gradient m e two directi

the image a r’s contour fore, using

our.

ernels of So

(a Fig. 14 The

has been an different me ch focuses o cribed in Fig magnitude in ions inform

g according to has appear

region gro

obel edge-d

a) e image (a) b

1 2 -1 -2

S1

+

15

n important ethods to so on gradient

g. 13. The v n the horiz

ation can de g=

o the magni red already owing segm

detection. T

before (b) a 2 1

2 -1

1

+

t issue in im olve it [18,

magnitude value calcu zontal direc erive the ov

itude g valu but there a mentation is

The sign “+

after the edg

-1 -2 -1

S2

+

mage proces 30]. In thes is generally lated by the tion and so verall gradie

ue and the re are still oth

s necessary

+” means th

(b)

ge detection 1

2 1

ssing for a se methods y known. A e first mask o does vert ent magnitu ( esult is like her edges in y to obtain

he center of

n.

long , the pair k, S1, tical.

de:

3.4) Fig.

n the n the

f the

(28)

B.

simp grow

whe dev from con grow com segm

Fig

C.

the subs are acco

Region Gro After the ple segmen wing metho

ere I(x) is th iation of th m each see dition in (3 wing step d mposed of a mentation m

g. 15 The reg

Seed Predic In dynam same and th sequent slic

correct and ording to th

owing Segm e pre-proces

ntation met od is adopte

he gray leve he initial reg d respectiv 3.5), those p does recursi

all seeds is method in th

gion growin

ction mic Elastogr

he seeds sel ces. Hence, d then the re he tumor di

mentation ssing and ed thod is eno d in the pap

el value of o gion and f i vely, if the

pixels beco ively until t the region he proposed

ng result of

raphy image lected by th

a seed pred esult of segm

splacement

16

dge detectio ough to fin per [28, 33].

original ima s a weight

neighborin ome new se there are no n of interes d system as

Fig. 14-(b)

es, the tumo he physician diction is ne mentation w in neighbo

on, the tumo nd a better . The formu

, age, m and

decided by ng pixels’ in eeds and the o new seed st (ROI). Th

illustrated i

, the white r

or position n in the first eeded to ens will be better oring slices.

or contour i result. Hen ula of region

σ are mea user. The p ntensity val e procedure ds appearing his region

n Fig. 15.

region is the

in each slic t slice could sure the see

r. The seeds However,

s obvious a nce, the re n growing is

( an and stan procedure s lues satisfy e goes on.

g. Then, the is the resu

e tumor reg

ce would no d not be use eds in each

s will be mo before the

and a gion s:

3.5) ndard

starts y the

This e set lt of

gion.

ot be ed in slice oved seed

(29)

17

prediction, because the tumor contour for the current slice i is not segmented out yet, the tumor displacement of the current slice can only be predicted by the displacement of the previous slice i-1. Moreover, the displacement of tumor centroid in neighboring slices is used to be the tumor displacement as illustrated in Fig. 16.

Fig. 16 The displacement of two slices i-2 and i-1 is used to predict the position of the tumor in slice i. The black point is the centroid of the tumor and the displacement is shown in dash line.

3.2 Elastography Image Quality Quantification

Conventionally, the physician selects a representative image from a sequence of dynamic elastography images dependent on his own subjective perception [11]. This paper proposes an objective quantification to measure the image quality for dynamic elastography images. In this section, two quantification methods SNRe and CNRe [12]

are introduced and the images with maximum quantification values will be selected as the representative images for the dynamic Elastography images.

3.2.1 Elastography Signal to Noise Ratio (SNRe)

In image processing techniques, the signal to noise ratio (SNR) [18] which represents the image quality is used generally. As its general formula:

SNR

^S

n (3.6)

where the s and n represents the magnitude of signal and noise, respectively. It is easy Slice i-2

Slice i-1

Slice i

(30)

18

to realize that SNR value is high while the signal is strong or the noise is less. In medical images, the SNR of elastography image shows the quality of the elasticity information and the formula is modified as (3.7) [12-14, 36, 37].

SNR

σ (3.7)

where s is the mean and σ is the standard deviation of the estimated strain and the measured region is shown as Fig. 17. A 10×10 tumor block T surrounding the centroid of the tumor is the region for computing SNRe. Comparing to (3.6), the mean of estimated strain represents the signal magnitude and the standard deviation represents the noise magnitude of the image. Because the mean of those images from the same movie file are similar, the SNRe value is high only when the standard deviation σ is small. That is, the estimated strain of the region is uniform and it is more probable to obtain the accurate strain information. Hence, the SNRe value could be correlated to the quality of elastography images. Examples of high and low quality images are shown in Fig. 18.

Fig. 17 The regions to quantify of elastography quality. The solid 10×10 rectangle is used to compute the SNRe and the all three rectangles are used to compute the CNRe. The x1 is the distance from the image border to the left boundary of tumor and the x2 is the one from the right boundary of tumor to image border. The distance from image border to the top boundary of tumor is represented by y. The upper-left corners of (x1/2)×y and (x2/2)×y background blocks B1 and B2 are located at (x1/4, y/2) and (M-3×x1/4, y/2), respectively,

y/2 y

Tumor

x1

x1/4

x1/2 y

x2

y

x2/4 x2/2 10

10

y/2

B1 B2

T

(31)

19

where M is the image width.

(a) (b)

Fig. 18 Two slices in a dynamic elastographic image. (a) The slice with low SNRe

value = 48.9362 and (b) The slice with high SNRe value = 153. 6835. The white rectangle represents the calculation region of SNRe. Note that the SNRe values are calculated from the hue information of the elastographic slices.

3.2.2 Elastography Contrast to Noise Ratio (CNRe)

Instead of focusing on the local uniformity like the SNRe, the contrast to noise ratio of elastography (CNRe), another quantification of elastography quality pays more attention to the contrast information of the tumor and background regions [12, 13]. The CNRe is defined as

CNR (3.8)

where the st and sb are the mean of the tumor region and the background region as shown in Fig. 17. The σt and σb are the standard deviations of the tumor region and background, respectively. The st and σt is obtained from the tumor block and the sb

and σb is computing from the background blocks B1 and B2 in Fig. 17. The B1 is a rectangle located on the upper left of the tumor and the B2 is located on the upper right of the tumor. The CNRe is useful in the low quality cases which the estimated strains of background are too similar to the strains of tumor like Fig. 19 (a).

(32)

Fig.

3.3

com

whe pixe pixe The in n

Fig.

. 19 Two sl value the ca from t

Hard Ratio In order mputed as th

ere the hard els in the tu el is over 20 e threshold o next chapter

. 20 The rel 200.

Hue

lices in a d

= 0 and (b) alculation re the hue info

o

to recogn he probabilit

d_num and umor. In thi 00, which is of the hard r.

lationship b e value 0 soft

(a) dynamic ela ) The slice egion of CN ormation of

ize the tum ty of being

H-r the total_n is paper, the s the thresho ratio is deci

etween hue

20

astographic with CNRe

NRe. Note the elastogr

mor as ben malignant.

ratio= ^__{_} num are the

e hard pixe old for the h ided accord

e value and

(b image. (a)

e value = 12 that the CN raphic slice

nign or ma The hard ra

e number of l is defined high or low ding to the e

strain. The 200

b)

The slice w 2.2662. The NRe values s.

alignant, the atio is defin

f the hard p d as if the h

strain as sh experiment r

threshold o 3 h

with low C e rectangles s are calcul

e hard rati ed as

( pixels and hue value of hown in Fig

result descr

of strain is a 360

hard

CNRe

s are lated

io is

3.9) total f the g. 20.

ribed

about

(33)

21

Chapter 4 Experiment Result

In the experiment of this paper, 141 biopsy-proved cases composed of 93 benign and 48 malignant lesions are used to evaluate the performance of the quality quantification. The benign cases include 54 fibrocystic changes, 12 papillomas, and 27 fibroadenomas. The malignant lesions include infiltrating carcinomas in 38 cases and ductal carcinoma in situ (DCIS) in 10 cases.

There are two experiments, the comparison of the hard ratios of different image quality quantification methods and the performance of two fast-selection methods. All the tumor contours used in these two experiments are obtained by the segmentation method stated in the previous section. In the first experiment, there are five ratios calculated by the proposed system for each case including four hard ratios of the slices with maximum SNRe, maximum CNRe, maximum compression, and maximum hard ratio and the hard ratio of the physician-selected slice. The hard ratio is calculated as described in the section 3.3. Each ratio can be used to differentiate the benign and the malignant cases and some statistic indexes are also used to evaluate the effectiveness of each ratio. In the second experiment, a fast-selection method is proposed to quickly select a slice with the best quality to calculate the hard ratio and the performance of this method is also evaluated by statistic indexes. Note that in the first experiment, all the tumors of slices in a movie file are needed to be segmented but in the second experiment only the tumor of the selected slice is needed to be segmented. Hence, the diagnosis time of fast-selection method could be much reduced.

(34)

22

The proposed system is implemented by using the programming language C++

under the Microsoft Visual C++ 2005 (Microsoft, Redmond, WA, USA). The operating system of the proposed system is Microsoft Windows XP (Microsoft, Redmond, WA, USA) on the Intel Pentium (2.66G dual-core machine with 3 GB RAM). The statistic analysis is performed with the use of the related software (SPSS, version 16 for Windows; SPSS, Chicago, IL, USA).

4.1 Quantification Performance

In the first experiment, there are five elastographic quality quantification methods needed to be evaluated. The hard ratio distribution of each method is described by the population pyramid [40], as shown in Fig. 22. The graphs of population pyramid show that the benign and the malignant lesions could be differentiated by the hard ratios, and moreover, the method using the hard ratios of the slices selected by maximum SNRe has the better performance for its more separability.

Assume that the hard ratios of the slices selected by the maximum hard ratio, SNRe, CNRe, compression are represented by Rmax, Rcnr, Rsnr, Rcomp, respectively and Rphy is for the ratios of the slices selected by physician. To ensure the ratios are useful, a statistical test is used. Because all the ratios in this experiment are the non-normal distribution, a Mann-Whitney U test [40] is performed and the result is listed in Table 1. Now, how to select the slice with maximum compression is explained. The central point of the tumor is used to evaluate the compression displacement. The vertical coordinate of the central point ymid is defined as:

(4.1) where Ttop and Tbottom means the vertical coordinate y’s of the highest and the lowest

(35)

23

apexes of the tumor. Because the larger compression will make the tumor position lower, the coordinate of central point could be used to select the slice with maximum compression. As illustrated in Fig. 21, the distance d of each slice is computed and the slice with the maximum d value is the slice with the maximum compression.

According to the result shown in Table 1, all the p-values of the Mann-Whitney U test are less than 0.05 which means the ratios of all quantification methods are statistical significant. Therefore, all hard ratios might be used to differentiate the benign and the malignant cases.

Even if all the hard ratios are proven to be useful to differentiate the benignancy and malignance, the differences between the ratios obtained by the proposed methods in this paper and the ratios obtained simply by the maximum ratio or by the physician are still unobvious. To examine whether the ratios proposed in this paper are better than Rmax, Rcomp, and Rphy or not, five performance indexes are introduced. These indexes are accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The performance indexes of the different hard ratios are shown in Table 2. The result shows that the ratio Rsnr and Rcnr are better than others, including the physician-selected hard ratio. Especially the method Rsnr has the best performance for its better sensitivity.

The receiver operation characteristic (ROC) curve is also used to evaluate the performance of the ratios. The ROC curves of the different ratios are shown in Fig. 23.

The area under curve values Az of Rsnr and Rcnr are also better than others, like the result of the statistical indexes above.

(36)

24

Fig. 21 An example to calculate the compression displacement. The distance d is equal to the compression displacement.

(a) Minimum ymid slice

d

A slice with larger compression

(37)

25

(b)

(c)

(38)

26

(d)

(e)

Fig. 22 The population pyramid of the hard ratio distribution in (a) Rmax (b) Rsnr (c) Rcnr (d) Rcomp (e) Rphy.

(39)

27

Table 1 The median value and p-value (using Mann-Whitney U test) of each hard ratio. All the ratios are non-normal distribution using the Kolmogorov-Smirnov test.

Ratio Type Median Value p-value

Rmax

Benign 0.2342

<0.001 Malignant 0.7598

Rsnr

Benign 0.0626

<0.001 Malignant 0.6174

Rcnr

Benign 0.0915

<0.001 Malignant 0.5881

Rcomp

Benign 0.0395

<0.001 Malignant 0.5321

Rphy

Benign 0.0708

<0.001 Malignant 0.5038

(40)

28

Table 2 Performance indexes of the proposed hard ratios.

Rmax Rsnr Rcnr Rcomp Rphy

Threshold 0.5177 0.3627 0.3865 0.1925 0.2752

TP 38 40 38 38 37 FN 10 8 10 10 11 TN 74 79 78 72 72

FP 19 14 15 21 21

Accuracy 79.43% 84.40% 82.27% 78.01% 77.30%

Sensitivity 79.17% 83.33% 79.17% 79.17% 77.08%

Specificity 79.57% 84.95% 83.87% 77.42% 77.42%

PPV 66.67% 74.07% 71.70% 64.41% 63.79%

NPV 88.10% 90.80% 88.64% 87.80% 86.75%

Note that:

Accuracy = (TP+TN) / (TP+TN+FP+FN) Sensitivity = TP / (TP+FN)

Specificity = TN / (TN+FP)

Positive Predictive Value (PPV) = TP / (TP+FP) Negative Predictive Value (NPV) = TN / (TN+FN) where

TP: Number of the positive cases true classified as positive.

FP: Number of the negative cases falsely classified as positive.

TN: Number of the negative cases true classified as negative.

FN: Number of the positive cases falsely classified as negative.

(41)

29

Fig. 23 The ROC curves of all hard ratios obtained by five methods.

4.2 The Fast-Selection Method

All the tumors in slices of the movie file have to be segmented to calculate the hard ratios in the first experiment. However, the segmentation task wastes a lot of time even if the segmentation method used in the proposed system is quite simple.

Indeed, the tumor contour is not really necessary for the computation of the SNRe and the CNRe if the compression displacement is not large. A fast-selection method uses some user-selected points, one for SNRe and three for CNRe, to define the tumor and background regions, as shown in Fig. 24. The central user-selected point is used to define the center of 10×10 tumor box and the upper-left and lower-right user-selected points are used to define the bounding box for the tumor in Fig. 17 for quantifying the

(42)

30

elastographic quality. After selecting the best quantified quality slices with maximum SNRe and CNRe, their hard ratios are computed. These ratios are represented by Rfsnr

and Rfcnr, respectively. The population pyramids of Rfsnr and Rfcnr are shown in Fig.

25.

To examine the effectiveness of the Rfsnr and Rfcnr, the same statistical evaluations to the first experiment are performed. The median value and the p-value of the Mann-Whitney U test are listed in Table 3. Both the Rfsnr and Rfcnr are statistical significant. The performance indexes are also used to evaluate the Rfsnr and Rfcnr and the result is listed in Table 4. The ROC curves are shown in Fig. 26 and the Az value of the Rfsnr is 0.8287 and the Az value of Rfcnr is 0.8451. Then, to find the relationship between the general quantification and the fast-selection method, the p-value of the ROC curves and the performance indexes are calculated. As illustrated in Table 5 and Table 6, the result of Rfcnr is more similar to the original ratios Rsnr and Rcnr than Rfsnr

in this experiment.

The performance of the ratios used the fast-selection method are worse than the origin Rsnr and Rcnr. The reason is that the fast-selection method uses just the user-selected points to derive the tumor region. However, the positions of the tumor are changing in different slices in a movie file, using the same tumor region will lead the decrease of the accuracy. Another point should be noticed is the Rfcnr is better than Rfsnr in both ROC curve and the performance indexes. The reason is that the fast-selection using SNRe refers to the region generated by the middle user-selected point only. This could make the Rfsnr be easier to be affected by the improper compression. Oppositely, the fast-selection using CNRe refers to not only the tumor region, but also the background region. The background region is much boarder than the tumor region so the effect of the compression could be reduced. Hence, the

(43)

31

fast-selection method using the CNRe is more appropriate to replace the general quantification methods than the fast-selection using SNRe. Note that the performance of the Rfcnr is very close to the performance of the original quantification methods with tumor segmentation of each slice and its run time is 58 second per case, which is the one-seventh of the run time of the original quantification method.

Fig. 24 The user-selected points ● of the fast-selection method. The central point is used in SNRe and all the three points are used to define the tumor and background regions in CNRe.

(a)

(44)

32

(b)

Fig. 25 The population pyramid of the (a) Rfsnr and (b) Rfcnr.

(45)

33

Table 3 The median value and p-value (using Mann-Whitney U test) of Rfsnr and Rfcnr.

Ratio Type Median Value p-value

Rfsnr

Benign 0.0531

<0.001 Malignant 0.5745

Rfcnr

Benign 0.0701

<0.001 Malignant 0.5948

Table 4 The performance indexes of Rfsnr and Rfcnr. Rfsnr Rfcnr

Threshold 0.2983 0.3518 TP 38 38 FN 10 10 TN 71 78 FP 22 15

Accuracy 77.30% 82.27%

Sensitivity 79.17% 79.17%

Specificity 76.34% 83.87%

PPV 63.33% 71.70%

NPV 87.65% 88.64%

(46)

34

Fig. 26 The ROC curves of the ratios using fast-selection method.

Table 5 The p-values of the ROC curves between the original quantification methods and the fast-selection methods.

Hard Ratios p-value

Rsnr Rfsnr 0.0042

Rsnr Rfcnr 0.0255

Rcnr Rfsnr 0.0850

Rcnr Rfcnr 0.2028

(47)

Tab

p Rsn

Rsn

Rc

4.3

by t hard

the few diag hard

le 6 The p- metho p-value

nr Rfsnr nr Rfcnr nr Rfsnr nr Rfcnr

Experim

In this se the quantific d to be diag

In Fig. 27 benignancy w cases like

gnosed corr d to differen

-values of th ods and the f

Accuracy 0.1302 0.6317 0.2993 1.0000

ment Resu

ection, som cation meth gnosed is als 7 and Fig. 2 y is extreme e the cases rectly only ntiate the be

(a)

he perform fast-selectio y Sensiti

0.601 0.601 1.000 1.000

lts

me experime hods of each

so explained 29, the lesio

ely low and stated in F by the elas enignancy a

35

ance indexe on methods ivity Spe

10 0

00 0

00 1

ent results w h case are in d.

ons are easy d in the ma

Fig. 28 an sticity infor and the mali

es between .

ecificity .1376 .8398 .1985 .0000

will be illus ntroduced a

to classify alignancy is d Fig. 30, rmation. In ignancy by t

the origina

PPV 0.2180 0.7822 0.3444 1.0000

strated. The and why the

because the s high. How which are these diffi the elastogr

(b)

al quantifica

NPV 0.5095 0.6369 0.8435 1.0000

e slices sele e case is eas

e strain ratio wever, there

difficult to cult cases, raphy.

ation

5 9 5 0

ected sy or

os in e are o be it is

(48)

36

(c) (d)

(e) (f)

(g) (h)

(i)

(49)

37

(j)

Fig. 27 A true negative example. (a)(b) The elastographic image and the B-mode image with the segmentation result of the same slice. (c) The slice selected by maximum hard ratio. The hard ratio of this slice is 12.24%. (d) The slice selected by maximum SNRe, 28.77. The hard ratio of this slice is 3.36%. (e) The slice selected by maximum CNRe, 13.18. The hard ratio of this slice is 3.44%. (f) The slice selected by the maximum compression and the hard ratio of the slice is 3.44%. (g) The physician-selected image with the hard ratio 3.29%. (h) The slice selected by the fast-selection using SNRe. The hard ratio is 0.58%. (i) The slice selected by the fast-selection using CNRe. The hard ratio is 12.85%. (j) The curves of different quantification methods. Each value of the quantification is normalized with the maximum value.

(50)

(a)

(c)

(e)

38

(b)

(d)

(f)

(51)

39

(g) (h)

(i)

(j)

Fig. 28 The false positive example. (a)(b) The elastographic image and the B-mode image with the segmentation result of the same slice. (c) The slice selected by maximum hard ratio. The hard ratio of this slice is 84.59%. (d) The slice selected by maximum SNRe,70.08. The hard ratio of this slice is 69.32%. (e) The slice selected by maximum CNRe,12.57. The hard ratio of this slice is 61.31%. (f) The slice selected by the maximum compression and the hard ratio of the slice is 34.08 %. (g) The physician-selected image with the hard ratio 65.83%. (h) The slice selected by the fast-selection using SNRe. The hard ratio is 63.66%. (i) The slice selected by the fast-selection using CNRe. The hard ratio is 70.61%. (j) The curves of different quantification methods.

Each value of the quantification is normalized with the maximum value.

(52)

(a)

(c)

(e)

40

(b)

(d)

(f)

(53)

41

(g) (h)

(i)

(j)

Fig. 29 A true positive example. (a)(b) The elastographic image and the B-mode image with the segmentation result of the same slice. (c) The slice selected by maximum hard ratio. The hard ratio of this slice is 82.45%. (d) The slice selected by maximum SNRe,102.59. The hard ratio of this slice is 82.06%. (e) The slice selected by maximum CNRe, 10.60. The hard ratio of this slice is 66.3%. (f) The slice selected by the maximum compression and the hard ratio of the slice is 68.25%. (g) The physician-selected image with the hard ratio 69.75%. (h) The slice selected by the fast-selection using SNRe. The hard ratio is 72.12%. (i) The slice selected by the fast-selection using CNRe. The hard ratio is 63.93%. (j) The curves of different quantification methods. Each value of the quantification is normalized with the maximum value.

(54)

(a)

(c)

(e)

(g)

42

(b)

(d)

(f)

(h)

(55)

43

(i)

(j)

Fig. 30 A false negative example. (a)(b) The elastographic image and the B-mode image with the segmentation result of the same slice. (c) The slice selected by maximum hard ratio. The hard ratio of this slice is 19.29%. (d) The slice selected by maximum SNRe,120.78. The hard ratio of this slice is 13.69%. (e) The slice selected by maximum CNRe, 14.88. The hard ratio of this slice is 14.3%. (f) The slice selected by the maximum compression and the hard ratio of the slice is 14.70%. (g) The physician-selected image with the hard ratio 0%. (h) The slice selected by the fast-selection using SNRe. The hard ratio is 5.55%. (i) The slice selected by the fast-selection using CNRe. The hard ratio is 11.49%. (j) The curves of different quantification methods. Each value of the quantification is normalized with the maximum value.

(56)

44

Chapter 5 Conclusion and Future Works

The elastography is generally used to display the elasticity information of the breast tissues now. In order to grade the lesion according to the elastographic image, Ueno [11] proposed a useful scoring system. However, most of the current methods to differentiate the benignancy and the malignancy are designed for only the still images.

For the convenience of diagnosis, a slice has to be selected from a movie file. The representative slice of a movie file is selected by the physician’s own subjective perception generally. To reduce the observer variability, this paper proposed a method to quantitatively select the representative slice which can represent the strain information of the whole data. With the seeds selected by the user, an automatic segmentation is performed first and then the quality quantification SNRe and CNRe

are computed. The experiment to evaluate the performance shows the higher accuracy of the maximum SNRe slices (84.40%, 119/141) and the maximum CNRe slices (82.27%, 116/141) compared to the maximum compression slices (79.43%, 112/141) and the physician selected slices (77.30%, 109/141). Although the performances of the slices selected by the proposed quantification methods are better, the run time of the proposed system is longer. The reason is that the tumor contours used in the quantification methods have to be segmented out in every slice. To save the executing time, the fast-selection methods using SNRe and CNRe are also proposed and only the tumor on first slice is needed to be segmented. The fast-selection method using CNRe

could also achieve a good performance with accuracy rate of 82.27% (116/141) in one-seventh time of the origin quantification method. This result indicates that the

(57)

45

proposed system can select a representative slice with an acceptable performance to differentiate the benignancy and the malignancy. With the use of the fast-selection method can further reduce the run time of the procedure because the segmentation is applied for only one slice, not all slices in a movie file.

The proposed method can automatically select a representative slice for the elastography. However, the method in this paper refers only the elasticity information and is lack of the help from conventional B-mode information. In fact, only utilizing the elasticity features could not diagnose the lesion correctly every time. According to Cho et al. [41, 42], the combination of the elastography and the conventional US could improve the sensitivity. Therefore, combining the quality quantification of the B-mode image and introducing the B-mode features to diagnose might obtain a better performance.