Press "Enter" to skip to content

Lung cancer detection using soft tissue technique in x ray image

Lung cancer detection using soft tissue technique in x ray image

Dr. R. Suresh Babu, Mr.R.Kishore

Haven't found the right essay?
Get an expert to write you the one you need
GET YOUR PAPER NOW

Department of ECE”,

Kamaraj College of Engineering and Technology”,

hodece@kamarajengg.edu.in and kishoreresearchscholar@gmail.com

Abstract

Lung cancer detection is a difficult task for radiologist since x ray image contain ribs and clavicles feature. It leads increase of FPs in existing CAD system. In order to reduce such FPs, single massive training artificial neural network is required for the creation of soft tissue technique from x ray image. It results in soft tissue image by which feature of x-ray image is being suppressed which proceed with segmentation. This work focuses on extracting 30 shape, gray level, texture and morphological features from each segmented candidate from both x-ray and soft tissue image. This proceeds with classification of candidates for the detection of lung nodules by the use of non linear support vector classifier. A publicly available JSRT database containing 247 images out of 154 x rays with nodules and 93 normal x rays was used to test this CAD system. All nodules were confirmed by computed tomography image and the average size of the nodule was 17.8 mm. By the use of the soft tissue technique, more nodules overlap with ribs and clavicles were detected and the sensitivity was 94% (145/154) with 2.5 FPs/image. Therefore, in x ray image, the number of FPs per image were reduced to 1.5 by reducing the values of sensitivity in proposed CAD system.

Keywords: lung cancer, sensitivity, specificity, classification, x-ray

Abbreviation used: CAD”,JSRT”,FP

Acronym used: Computer Aided Detection, Japanese Society of Radiological Technology, False Positive

Keypoints:

i) Early detection of nodule candidate increases the survival rate

ii) Sensitivity of a CAD system is directly proportional to true positive rate

iii) Specificity of a CAD system is inversely proportional to false positive rate

I. Introduction

In every year (2009), American Cancer Society announced the overall 5-year survival rate of lung cancer patients. It is only 14% for all combined stages (early, middle & advanced stages) [1]. Nodule candidate detection in soft tissue image can improve the survival rate by 50%. In recent days, it is a challenging task for radiologist due to the suppression of nodule feature.

A sensitive imaging modality was needed for lung cancer detection [2]. X-ray image is being preferred because this is cost-effective, routinely available and dose-effective diagnostic tool [3]. But, 30% of nodules in x ray image were missed by radiologists and that, 82-95% of missed nodules were partly obscured by overlying bones such as ribs and clavicles [4][5]. This image suit to improve the step on nodule candidate detection.

A major challenge for current CAD system is to detect the nodules overlapping ribs, rib crossings, and clavicles because one-third of the nodules were missed by radiologists. This leads to lower the sensitivity. Feng Li, Roger Engelmann, Kunio Doi, Heber MacMahon detect small lung cancers by the use of dual energy subtraction chest radiography to assist radiologists in their detection and increase radiologist efficiency [6]. In addition, radiologists may increase their confidence with the CAD system as a useful tool. In recent days, the above mentioned

technique is applicable for a limited number of hospital due to the requirement of specialized equipments with double

dose radiation. Matsumoto et.al., proposed the CAD system at a FP rate of 11 per image, even though the system had a high sensitivity of 80%. But the accuracy of detecting lung nodules was not improved [7].

In this work, we propose a CAD system for the detection of lung nodules by the use of soft tissue technique to improve the sensitivity with the same FP rate and simultaneously reduces the FP rate by reducing the value of improved sensitivity. Hence, the accuracy of this CAD system was improved for radiologists to make a good decision.

II. Materials and Methods

A. Database of X-Ray Image

In this proposed CAD system, we collected 154 images with nodules and 93 normal images from the publicly available JSRT (Japanese Society of Radiological Technology) database [8]. The nodule size ranged from 15 to 24 mm. The images were digitized to yield 12-bit chest x rays (CXRs) with a resolution of 2048 × 2048 pixels. The pixel size was 0.175 × 0.175 mm. Missed nodules were divided into five categories based on subtlety rating. Among subtlety rating, other information such as age, gender, nodule size, malignancy status, coarse anatomic location and co-ordinates of the nodule center. The five categories are extremely subtle, very subtle, subtle”,

relatively obvious and obvious.

B. Existing CAD System

The images were collected from the University of Chicago Hospitals in the Department of Radiology and were obtained by the use of single exposure technique with two detectors seperated by a copper filter [6]. The dimension of the original image are 1760 x 1760.The images are linearly sub-sampled to 512 x 512 dimension. This was done to facilitate experimentation as it allows for a considerable decrease in computation time. In order to segment lung, a various segmentation methods was used to extract lung boundary [9]. The various segmentation methods are rule based methods, pixel based methods, hybrid methods and deformable model based methods.

Other essay:   About key details one need know concerning cervical cancer

After lung fields were segmented, a background trend correction technique was applied to the segmented lung field. A second order bivariate polynomial function was fitted to each of the left and right lung fields F(u”,v)= au2+bv2+cuv+du+ev+f where a, b, c, d, e and f are co-efficient calculated for each parameter. Subsequently, the fitted functions for the left and right lung fields were subtracted from the original image a(u”,v). In this approach, a background trend in the lung fields as a low frequency surface was removed [10].

The pre-processed image in existing CAD system was applied to different types of gray level morphological open operation to form a nodule enhanced image and a nodule enhanced image is modified in the nodule likelihood map [11]. It results in the detection of nodule candidates based on local peaks in the map. This nodule candidate is very important to find the lung nodules which is used to locate the boundary of the nodule.

Peaks within the nodule candidate region in the nodule enhanced image is used to initialize the candidate segmentation by the use of watershed technique. With the watershed technique [12], the nodule candidate region was

divided into several gray levels. The advantage of this technique is able to produce good segmentation results and it would be more useful to focus on nodule candidates. Finally, gray level features were extracted from nodule candidates.

The name of the gray level features are contrast, correlation, energy, homogeneity and entropy. Based on the extraction of gray level features, feature selection method based on genetic algorithm [13] is used to determine the optimal size of a feature set automatically. The optimal size of a feature set is given as input to support vector machine classifier for classifying nodule candidates into nodules or non-nodules [14].

A major issue for the detection of lung cancer in existing

CAD system is to detect lung nodules with overlapping ribs

and clavicles in pre-processed image. Due to this issue, it leads increase of false positives in existing CAD system. So, the sensitivity of CAD system is also reduced. In order

to improve the sensitivity, we propose a soft tissue

[image: image1.png]

Fig 1 Existing CAD system

technique to detect nodule candidate in soft tissue image. So, the false positive of proposed CAD system is also reduced.

C. Proposed CAD System

A soft tissue technique is a technique used to suppress ribs and clavicles in x ray image with the help of massive training artificial neural network (MTANN). With this technique”,feature of the nodule candidate are being suppressed. Fig 2 shows the creation of soft tissue technique for the detection of nodule candidate by the use of MTANN.

[image: image2.png]

Fig 2 Proposed CAD system

MTANN is a highly non linear filter which needs to be trained by the use of x ray image and corresponding teaching image. Bone image from dual energy radiography system were taken as the teaching image.

fb(u”,v) =NN (au”,v) ———————————————-(1)

Where au”,v = {g(u-i”,v-j) |i”,j є Rs } is the input vector to the MTANN which is a subregion extracted from x ray image and fb(u”,v) is an estimate for a teaching value.

The MTANN is modified by using the subregions together with each of the corresponding teaching single pixels. The

training set of pairs of a sub region and a teaching pixel is

represented by

{a(u”,v), T(i”,j)| u”,v є RT }

= { (a1 “,T1) (a2″,T2) (aN”,TN) }————————–(2)

Where T(i”,j) is a teaching image, RT is a training region which corresponds to the collection of the centers of sub region and N is the number of pixels in RT.

For a single MTANN, suppression of ribs containing various frequencies was difficult due to limited capability. In order to overcome this issue, multiresolution decomposition/composition techniques were applied. First, a lower resolution image GL(u”,v) was obtained from an original higher resolution image GH(2u”,2v) by performing downsampling with averaging i.e., four pixels in the original image are replaced by a mean value of four pixels represented by

GL(u”,v) = (1/4) ∑ GH(2u, 2v)—————————(3)

u”,v є R22

where R22 is a 2 x 2 pixel region. The lower resolution image is enlarged by upsampling with pixel substitution i.e., a pixel of lower resolution image is replaced by four pixels with the same pixel value, as follows:

GU(u”,v)=GL(u/2″,v/2)——————————————-(4)

Then, we subtract enlarged lower resolution image from the higher resolution image to get high resolution difference image, represented by

DH(u”,v) = GH(u”,v) – GU(u”,v) ——————————–(5)

These procedures are performed again and again producing further lower resolution images. Thus, multiresolution images with various frequencies are obtained by the use of multiresolution decomposition technique. An important property of this technique is the same high resolution image GH(u”,v) that can be obtained as follows:

Other essay:   Pet/ct in assessment of ovarian cancer recurrence

GH(u”,v)=GU(u”,v)+DH(u”,v) ————————————(6)

Therefore, we can choose multiresolution images for processing independently instead of processing original high resolution images directly. After training with input x ray image and the dual energy bone image, the multiresolution MTANN is able to produce bone image that are expected to be similar to the teaching bone image. The bone image fb(u”,v) produced by the trained MTANN with the lung masking image n(u”,v) and weighting

parameter wc is subtracted from the sub region of original x ray image g(u”,v) to create a soft tissue image.

f(u”,v)= g(u”,v) – wc x fb(u”,v) x n(u”,v)————————(7)

where f(u”,v) denotes the soft tissue image with different

types of rib contrast by the use of weighting parameter wc.

The proposed CAD system segments the lung for identifying sub region, create the soft tissue technique to

detect nodule candidate on soft tissue image, segments the nodule candidates on soft tissue image which proceed with

feature extraction and classification for the detection of lung nodules. Segmentation of nodule candidates is done

D. Existing and Proposed CAD System

[image: image3.png]

Fig 3 Existing and Proposed CAD system

by the use of clustering watershed method. Fig 3 shows the combination of existing and proposed CAD system for the detection of subtle nodules to reduce rib-induced false positives. Peaks within the nodule candidate region in the nodule enhanced image were obtained and used to initialize the clustering watershed method [12]. With the clustering watershed method, the nodule candidate regions were divided into several catchment basins. Each minimum point were surrounded by a catchment basin associated with it; there were one or more peaks, each of which were surrounded by a cluster of connected pixels that consituted a catchment basin.

The features of the nodules were suppressed in soft tissue

image. Detection of nodule candidate on soft tissue image

may be misclassified as non-nodules. So, the segmentation step is being repeated on soft tissue and x ray image. Sixty morphologic and gray level based features were extracted from each candidate from both x ray and soft tissue image and it is given as input to feature selection method based on genetic algorithm [13]. A non linear SVM with a Gaussian kernel was employed for the classification of nodule candidates into nodules or non-nodules [14].

III. Results

In this section, we present some results to demonstrate the performance of proposed CAD system. First, we create the soft tissue technique for the proposed CAD system. Next, the sensitivity for nodule candidate detection on soft tissue image with different rib contrast was plotted to find the peak value of nodule candidate.

A. MTANN Training

Four x rays and their corresponding bone images in a training set were used to train the MTANN. It is one of the advantages that it needs four training images. In the four images, one was a normal image and the other three contained with nodule image.In a sub region, the size of the MTANN was 9 x 9 pixels. It was sufficient to cover the width of rib in the image. Three layered MTANN were used to suppress rib where the number of input, hidden and output units were 81, 20 and 1 respectively.

B. Nodule Candidate Detection

[image: image4.png]

Fig 4. Effect of the change in rib contrast on the sensitvity of the proposed CAD system

In fig 4, the sensitivity was highest when the rib contrast

parameter mc was set 0.4. The most nodules were obvious in the nodule candidate detection step. In this plotted graph”,

84 candidates that had max code values in the nodule likelihood map were selected as the origin point to indicate

the performance of the nodule candidate detection

C. Performance of the existing CAD system using soft tissue image

Soft tissue image was applied to replace the x-ray image in existing CAD system. The nodule candidates were detected from the soft tissue image by the use of two stage nodule enhancement method. Fig 5 illustrates the performance improvement of nodule candidate detection step by the use

of soft tissue image. The non-subtle nodule overlapping ribs and clavicles missed in soft tissue image but detected in x-ray image. So, the performance of existing CADs system using x ray image has been improved to a minimum. Then, 30 features were extracted from the soft tissue image after the segmentation of nodule candidate.

[image: image5.png]

(a) (b)

Fig 5(a) Nodule candidate detection results based on x ray image (b) Nodule candidate detection results based on soft tissue image

[image: image6.png]

Fig 6 Sensitivity vs Number of FPs per image

Table I: Features Extracted for Proposed CAD system based on the soft tissue and x ray image

S.No

Feature Extraction

Description

1

can.u

Co-ordinates of a nodule candidate

2

can.v

Co-ordinates of a nodule candidate

3

can.Grad1

Co-ordinates of a nodule likelihood values

4

can.CV1

Calculated by using gray level values

5

can.Grad2

Co-ordinates of a nodule likelihood values

6

Can.CV2

Calculated by using gray level values

7

Shape1

Area of a segmented nodule candidate

8

Shape2

short and long axes of an ellipse fitted to the nodule candidate

Other essay:   Understanding the computational physics of cancer

9

Shape3

Area of a convex hull of the candidate

10

Shape4

distance between the centroid of the candidate and the centroid of the fitted ellipse

11

Gray1

μregion – μsurround

12

Gray2

σregion- σsurround

13

Gray3

minregion-minsurround

14

Gray4

maxregion-maxsurround

15

Gray5

Calculated using Gray1

16

Gray6

Calculated using Gray2

17

Gray7

Calculated using Gray3

18

Gray8

Calculated using Gray4

19

Grad1

7

= Gr = (1/8) ∑ Grh

k=0

Grh = [1/Nh ] ∑ cos αmn

mn є regionh

t1 ≤ Mmn ≤ t2

where Nh is the number of pixels in the segmented nodule region h

20

Grad2

7

= σ = √ ∑ (Grh- Gr )2

k=0

21

Grad3

= Gr / σ

22

Surface1

λmin

23

Surface2

λmax

24

Surface3

λmin λmax

25

Texture1

∑ [C(i”,j)2]

ij

26

Texture2

∑(i-j)2C(i”,j) ij

Where C(i”,j) ( Co-occurrence

matrix calculated over neighboring pixels and a summation range from the minimum to the maximum pixel value in the pre-processed image

27

Texture3

Calculated based on Texture1 and Texture2

28

Texture4

Calculated based on Texture1 and Texture2

29

Texture5

Calculated based on Texture1 and Texture2

30

Texture6

Calculated based on Texture1 and Texture2

31

False Positive

Lregion / Loverlap

Table II Performance comparison of several CAD systems which used the JSRT database

Sensitivity FPs/image Database

Wei et al. 80% 5.4 All

(123/154) (1333/247) nodule and

normal

image in

JSRT(247)

Coppini et al. 60% 4.3 All nodule

(93/154) (662/154) image in

JSRT(154)

Schilham et al. 67% 4 All nodule

(103/154) (616/154) image in

JSRT(154)

Hardie et.al. 63% 2 Nodule

(88/140) (280/140) image in

JSRT(140)

Chen et.al. 71% 2 Nodule

(100/140) (466/233) and Normal

image in

JSRT(233)

Soft tissue 87% 1.5 All nodule

based CAD (134/154) (371/247) and system normal

image in

JSRT(247)

IV. Conclusion

In a proposed CAD system, a soft tissue technique is a technique used to suppress ribs and clavicles with the help of massive training artificial neural network. Detection of nodule candidate on soft tissue image misclassified as non-nodules due to the suppression of nodule feature. So, it is being developed with sensitivity value calculated as 94%=145/154 by incorporating soft tissue technique after the segmentation of lung. The number of FPs per image was reduced to 1.5 by reducing the values of sensitivity to 87% in proposed CAD system.

The authors have declared that there is no conflict of interest.

References

[1] Arnold M.R. Schilham, Bram van Ginneken, Marco Loog (2006), ”A computer-aided diagnosis system for detection of lung nodules in chest radiographs with an evaluation on a public database”, Medical Image Analysis 10, pp 247–258.

[2] C.I. Henschke, D.I. McCauley, D.F. Yankelevitz, D.P. Naidich, G.McGuinness, O.S. Miettinen, D.M. Libby, M.W. Pasmantier, J. Koizumi, N. K. Altorki, and J. P. Smith(1999), Early lung cancer action project: Overall design and findings from baseline screening, Lancet, vol. 354, pp. 99 – 105.

[3] Chen S, Suzuki K, MacMahon H (2011), “Development and evaluation of a computer-aided diagnostic scheme for lung nodule detection in chest radiographs by means of two-stage nodule enhancement with support vector classification”, Medical Physics”,

Apr 38(4), pp 1844-58.

[4] J.H Austin, B. M. Romney, L.S Goldsmith (1992), Missed bronchogenic carcinoma: Radiographic findings in 27 patients with a potentially resectable leison evident in reterospect, Radiology, vol. 182, pp 115-122.

[5] P. K. Shah, J. H. Austin, C. S. White, P. Patel, L. B. Haramati, G. D. Pearson, M. C. Shiau, and Y. M. Berkmen (2003), Missed non- small cell lung cancer: Radiographic findings of potentially resectable lesions evident only in retrospect”, Radiology, vol. 226, pp. 235–241.

[6] Feng Li, Roger Engelmann, Kunio Doi, Heber MacMahon (2008), Improved detection of small lung cancers with dual-energy subtraction chest radiography, AJR Amer. J Roentgenol, vol.190, pp 886-891.

[7] T. Matsumoto, K. Doi, A. Kano, H. Nakamura, T. Nakanishi(1993), Evaluation of the potential benefit of computer aided diagnosis of lung cancer screenings using photofluorography: analysis of an observer study, Nihon Igaku Hoshasen Gakkai Zasshi, vol. 53, pp 1195-1207, Oct 25.

[8] http://db.jsrt.or.jp/eng.php

[9] William M.Wells, Allan Colchester, Scott Delp (1998), “Medical Image Computing and Computer Assisted Intervention”, Springer.

[10] Bram van Ginneken, Alejandro F. Frangi, Joes J. Staal, Bart M.ter Haar Romeny, Max A. Viergever, (2002), Active shape model segmentation with optimal features, IEEE Transactions on medical imaging, Vol.21, No.8.

[11] Haralick RM, Sternberg SR, Zhuang X(1987), Image analysis using mathematical morphology, IEEE Transactions on Pattern Analysis and Machine Intelligence, pp 532-550.

[12] Chongbo Zhou, Chuancai Liu, “Co-segmentation of multiple similar images using saliency detection and region merging”, Computer Vision IET, vol. 8, no. 3, pp. 254-261, 2014

[13] Lilla Boroczky, Luyin Zhao, K.P. Lee (2005)”,Feature subset selection for improving the performance of false positive reduction

in lung nodule CAD, IEEE symposium on computer based medical

systems.

[14] Sheng Chen, Liping Yao, Bao Chen (2016) A parameterized logarithmic image processing method with Laplacian of Gaussian filtering for lung nodule enhancement in chest radiographs. Medical & Biological Engineering & Computing 54:11, pp 1793-1806.

Be First to Comment

Leave a Reply

Your email address will not be published.

0 Shares
Share via
Copy link

Spelling error report

The following text will be sent to our editors: