1. Introduction

Image segmentation is one of the important tasks for image understanding and analysis. Segmentation partitions an image into a set of different regions. Each segmented object is homogeneous with respect to certain intensity characteristics like color and texture. The segmented objects in an image can be further separated from the background of the image. Most of the existing methods have used mostly thresholding concepts for segmentation. The different methods used for segmentation are statistical, geometrical and structural. Moreover, high-resolution XRay Computed Tomography (CT) is the standard for representing pulmonary images [10]. CT scans provide high spatio-temporal resolutions based on the hardware used for scanning. It also provides an outstanding contrast resolution for images pertaining to the pulmonary structure. Moreover, it has the capability to form a complete threedimensional (3-D) volume [23, 25] of human thorax for a single breath. In the diagnosis of lung cancer, pulmonary CT-scanned images have been used for many applications including lung parenchyma analysis, density analysis and nodule detection. Lung segmentation is a precursor to most quantitative analysis [24].

Segmentation of medical imaging is useful for extracting information about the status of different tissues, different organs and other parts of the body [13]. The segmentation accuracy and higher decision confidence value of any lung analysis system for the identification of abnormalities rely mostly on an efficient lung segmentation technique. Hence the performance of segmentation based systems depends on the lung images [22]. In this paper, we propose a Fuzzy Logic based Bitplane [26] algorithm (FLBP) for the extraction of lungs from the background of a CT lung image. The extracted lungs are further used for identifying nodules and to perform further analysis.

The salient features of the proposed (FLBP) technique are:

a. It is an unsupervised method and is completely automatic. b. No prior assumption is made about the type of features and type of contents, amount of noise and the contrast of the image. c. The proposed method has a novel segmentation process based on assignment of 0 or 1 to pixels. d. This method is robust in segmentation due to the incorporation of Fuzzy Logic with Bitplane. e. The segmentation is accurate since it considers foreground and background of the image separately.

This FLBP method ensures that no excess pixel along the border of lungs is segmented. The remainder of the paper is organized as follows: Section 2 provides a survey of related works. Section 3 explains the proposed FLBP algorithm in detail by providing the details of each step. Section 4 depicts the results obtained from the work. Section 5 presents the conclusions on this work and suggests possible future works.

 

2. Related Works

Segmentation of lung in a high resolution CT (HRCT) image is performed using a pixel-based approach was proposed by Garnavi et al [1]. In their work, the decisions were based on crisp rules. Threshold based segmentation approach for segmenting lung region was proposed by Lin DT et al [7]. They have used a 5×5 median filter for removing the noise. The Foreground is separated by omitting the rim of the image along with the background. Samuel et al [4] have proposed a Rolling Ball-Algorithm for the segmentation of lung region from the CT lung image. They have applied the gray-level thresholding to obtain the lung regions. Moreover, they have segmented the lung region initially by removing the thorax from the background and then the lung region from thorax. Shiying et al [6] have developed an automatic method for segmenting lungs in CT lung images. They have divided their work into three main stages: (1) Gray-level thresholding to extract lung from CT-Scan image (2) Left and right lungs are separated through the identification of the anterior and posterior junctions by using dynamic programming and (3) smoothing of the irregular boundaries along the mediastinum, a sequence of morphological operations for acquiring results.

A combination of background-removal operator and iterative gray level thresholding was proposed by Antonelli et al [2]. In their work, the background was not eliminated well due to the presence of noise. Adaptive border marching algorithm was proposed by Jiantao Pu et al [9] which segments the lung region and reduces the under segmentation. They used the gray-level thresholding to obtain the lung regions and a flood-filling methodology to remove non-lung regions present after the thresholding. Ozekes et al [11] segmented the lungs of the CT images using Genetic Cellular Neural Networks (G-CNN). In their work, the lung regions were specified using the 8 directional searches and +1 or −1 value were assigned to each voxel. In the work proposed by Cao Leiet al [8], a rough image of lung was acquired by a combination of optimal thresholding and mathematical morphology. A selffit segmentation algorithm was applied on the segmented result to obtain a refined output.

Manish Kakar [3] proposed a novel lung extraction algorithm by extracting the texture features using Gabor filter for filtering. In their algorithm, the features are combined to segment the lung region by using Fuzzy C Means clustering. Arfan Jaffar et al [5] proposed a combination of spatial Fuzzy C-Means and morphological techniques for segmenting the lung region from CT lung images. They have divided their work into three main stages namely Background removal, Preprocessing, and Morphological based operations.

Among them, a new histogram background removal operator has been used to remove the pixels around the lungs. Moreover, preprocessing is performed by smoothing and removing the noise present in the segmented image. Finally, morphological operations are used to separate the edges and to fill the small holes present in the segmented lungs. A joint serial image registration and segmentation methodology was proposed by Zhong Xue et al [16]. The major purpose of this algorithm is to segment a set of serial lung CT images. In their algorithm, the serial CT images are segmented based on the current deformations, and then the deformations among the serial images are iteratively refined based on the updated segmentation results. The main drawback of this technique is the presence of noise and discontinuities in the segmented lung region.

From the works found in the literature, it has been observed that most of the existing works used. Thresholding, Fuzzy C means, Neural networks and Background removal operators. However in case of medical applications, the accuracy provided by various phases of segmentation is not sufficient to make effective decisions. Therefore, it is necessary to propose a new and efficient technique to enhance the accuracy of segmentation.

 

3. Overview of the FLBP Method

In this paper, a fuzzy logic with bitplane method has been proposed for effective segmentation of lung images. This combination of fuzzy logic with bitplane method helps to make effective decisions under uncertainty. The FLBP method is a combination of bitplane processing of the image and fuzzy logic applied to each pixel. The FLBP method consists of the following steps:

a) The lung border values in the image are considered to decide the threshold value. b) The gray scale lung image is separated into 8 monochrome images and hence the bit values from each bitplanes are obtained. c) The 4 bits corresponding to the background of the intensity values are named as “lower nibble (LN)” and the other 4 bits corresponding to the foreground of the intensity values are considered as “upper Nibble (UN)”. d) Eight features are formed at each pixel. They are v8, v7, v6, v5, v4, v3, v2, v1. The numbers in the feature variables indicate the bitplane. e) To assign a segmentation value of 1 or 0 to a pixel, a window of 3 × 3 pixels is considered. The 3 × 3 window is moved horizontally and vertically to cover the entire image. f) Three signum membership functions are used to form a fuzzy membership value table in order to assign a bit value 0 or 1 to the bitplanes. g) Inference is carried out for the foreground and background based on the bit values assigned using fuzzy membership table. h) Defuzzification is performed with the maximum number 1 or 0 in a 3 × 3 windows considering foreground and background bit values of each pixel to decide the final segmentation. i) Lung region is separated from the original image.

3.1 Description of proposed methodology

The FLBP system is based on bitplane processing. In this paper, the term bitplane is used to show the details of the image at various depths of information. Fuzzy rules are developed and used for determining the optimum segmentation of the lung from the background of a CT image. In this work, CT lung images are taken and represented using Hounds Unit (HU). Gray scale CT lung image is the normalization of the Hounds unit (HU) of the original CT image [12].

Table 1 shows the approximate representation of HU and gray scale. We have taken gray scale image as each pixel can be analyzed at bitplanes using the proposed FLBP method. A gray scale value of (63.25) representing the lung region was shown in the above table. The value can be 63.25 ± 40 to cover the entire area of lungs. The basic block diagram of the proposed system is given in Fig. 1.

Table 1.Representation of Hounds Unit and gray scale

Fig. 1.Block diagram of FLBP system

3.1.1 Fuzzification of the Image

Fuzzification is the process of transforming the values from one crisp sets in to fuzzy set members for qualitative representation. In many Fuzzy segmentation methods like Fuzzy clustering, Fuzzy C-means and Fuzzy inference, transformation of intensity values to a different range of values is considered as an initial process. In this work, the transformation (Fuzzification) process is equivalent to the formation of 8 bitplanes and they are shown in Figs. 3(a-h). Each image contains 1or 0. Hence, each bitplane is a monochrome image. By observing each bitplane, the gradient of the boundary of lungs is calculated and a decision is made whether segmentation is required at such gradients. The degree of membership of each object in the image is computed on its association to each bitplane. The histogram of each bitplane shows a degree of association of each object in a bitplane.

Fig. 2(a) presents the two lungs which are clearly visible from the background in bitplane 8. Profiles of the two lungs are clearly visible in Fig. 2(b) which is present in bitplane 7. The bitplane 6 represents the clearly visible center portion of the CT lung image. In Figs. 2(d-h), the lungs are not clear and the information appears to be noisy.

Fig. 2Bitplane extraction results: (a) Bitplane-8; (b) Bitplane-7; (c) Bitplane-6; (d) Bitplane-5; (e) Bitplane-4; (f) Bitplane- 3; (g) Bitplane-2; (h) Bitplane-1.

3.1.2 Feature Extraction

Feature extraction is a process of extracting representative values that will uniquely distinguish one image from a group of images. In this work, features extracted from lungs are different from the features of other parts of lungs. Many methodologies are available for extracting features of various objects in the lung image. These features are usually normalized values. The 8 features obtained at each pixel are grouped into two categories namely foreground and background features. Features (V8, V7, V6, V5) belongs to the background and features (V4, V3, V2, V1) belonging to the foreground. During the segmentation process, a 3x3 window of pixels are considered at a given instance. Hence, 8 features at each pixel and features of all the 9 pixels are considered to assign bit value either 1 or 0 at the center of 3x3 window.

Table 2 presents the feature variables and its possible bit values that can be present at different bit level. The first column of Table shows the feature variables. The subsequent columns represent monochrome values present at different bitplane of a pixel. Depending upon the gray scale value, the value of each bitplane can be 1 or 0. The foreground of the gray scale is in bitplanes (8 to 5). These bitplanes are named as Upper Nibble (UN). Similarly, the background of the gray scale is in bitplanes (4 to 1) is named as Lower Nibble (LN).

Table 2.Feature variables and its Calculation of feature values considering different bitplanes

During the calculation of the foreground features, the lower bitplanes are replaced with 0 inorder to project the foreground of the image. Similarly, for the calculation of background features, the upper bitplanes are assigned with 0 inorder to project the foreground of the image. This can be noticed in Figs. 3(e-h) which show information in the background is not much useful for segmentation.

Fig. 3.Triple signum function

3.1.3 Membership Table Formation

Membership degree can be expressed by a mathematical function μA(x1) that assigns each element in the set, a membership degree between 0 and 1. Let X is the universe of discourse and xi is an element of X. A fuzzy set A in X can be defined as [14].

The signum function [19] (Sig) is used for modelling the membership degrees. This degree of function is suitable for representing the white and black pixels. In this proposed work, we are focusing on segmentation by using bits. In such case, we get either 0 or 1rather than getting any intermediate outputs. Hence, we are choosing only signum function. Two breakages are required in the signum function in order to segment the UN and LN separately.

In this paper, we propose a new signum function called triple signum function that outputs the sign of a real number. This function outputs any one value from the three values (−1, 0, 1), when a real number is given as input to the proposed triple signum function. In order to propose this function, the existing signum function has been modified in this work for suiting our problem by outputting the bit values (1, 0) and hence the outputs of the signum function are (1, 0). This function can be also called as step function since only (1, 0) are obtained. Three signum functions are used in this work since three intervals levels of thresholds are used.

The width of signum function is controlled by two variables ‘a’ and ‘b’. Similarly, the width of signum function 2 and 3 are controlled by ‘c’ which is present between ‘b’ and ‘d’. Depending upon the object to be detected, the location of ‘c’ can be changed.

This triple signum function as shown in Fig. 3 consists of three functions namely sig1, sig2, and sig3. Among them, sig1 and sig3 are the positive signum functions. As the values of the feature variables increase, the signum value changes from 0 to 1. Similarly in sig2, when the values of the feature variables increase the signum value changes from 1 to 0. Hence, sig2 is called as negative signum. For the UN bits, the sig2 and sig 3 are used.

The sig2 and sig3 function are suitable to represent the set of UN bits and is defined as

The sig1 and sig2 function is suitable to represent the set of LN bits and is defined as

The membership table is similar to a look up table and its formation is required to associate the input information in to output information. Membership table provided in Table 7 helps to assign the degree of membership value for the intensity values of different pixels.

3.1.4 Inference

The inference is a process of forming a new decision based on existing information. In this work, inference is carried out inorder to form the membership table by assigning output values based on feature values present in Table 6. At the end of inference, only two segmentation outputs are obtained at each pixel. The numbers in parenthesis in Table 7 are the bit values assigned based on the fuzzy rules shown below:

Table 6.Total intensity values calculation

Table 7.Membership table for the 9 intensity values based on Fuzzy rules

3.1.5 Defuzzification

Defuzzification is the process of combining the outputs of the inference for each pixel to a final segmented output that provides a segmented lung against the background of CT image. There are atleast 15 different defuzzification methods. Moreover, we have used “Maxima” defuzzification method [17]. Due to its generality, maxima refers to the total number of occurrence of either 1 or 0.The four steps used for defuzzification are as follows

The result obtained from final_segmentation procedure which was assigned to the 3*3 window is shown in Table 11. These values 0 and 1 help to get the final segmented lung region.

Table 11.Final segmentation values

 

4. Experimental Evaluation

Segmentation accuracy has been evaluated in this work to assess the performance of proposed algorithm. In this section, we present the results of the proposed method qualitatively and quantitatively through image display and experiment measurements respectively. If the segmented object does not have any other smaller objects when compared to the same location of the original image, then it’s easier to evaluate the segmentation accuracy based on counting the number of pixels in the original image.

4.1 Existing metrics to measure the segmentation accuracy

The Accuracy metrics As proposed by Zulaikha Beevi [18] is used to measure the accuracy of medical images. In our method, the segmentation accuracy As is computed using the formula.

where

Apixel= segmentation accuracy Nc = number of correctly segmented pixels in the lungs Tp = total number of pixels corresponding to the lungs in the unsegmented image (Ground Truth)

Four grayscale CT images are considered for evaluating the quality of segmentation for our proposed method.

4.2 Proposed metrics to measure the segmentation accuracy

We have proposed another method to evaluate the segmentation accuracy by using the region properties of Matlab. The important properties like solidity, perimeter and Area can be used.

The segmentation accuracy Aprop is computed using the following equation to evaluate the quality of the segmented lungs:

where Sprop and Oprop are the region properties of the segmented and unsegmented images respectively.

Solidity is a scalar specifying the proportion of the pixels in the convex hull that are also in the region [15]. Area is the actual number of pixels in the region of interest. Convex Area is a Scalar that specifies the number of pixels in Convex Image. Convex Image is a Binary image (logical) that specifies the convex hull, with all pixels within the hull filled in.

The image is the size of the bounding box of the region. This property is supported only for 2-D input label matrices. Perimeter is the distance around the boundary of the region. Region props compute the perimeter by calculating the distance between each adjoining pair of pixels around the border of the region

To prove the efficiency of the proposed segmentation algorithm, a sample 3×3 intensity values shown in Table 5 are taken from the border of the left lung from Fig. 4 (a) which are used in this paper to explain the segmentation process. The symbols p1, p2, 93, 94, 95, p6, p7, p8 and P9 indicate the pixel number.

Table 5.Sample intensity values of 3 × 3 window

Fig. 4Feature value representation

Table 6 shows the computed total intensity values of features V8, V7, V6, V5 for the foreground and V4, V3, V2, V1 for the background based on the sample 3x3 intensity values. The intensity values assigned to each variables are based on the logic in Tables 3 and 4.

Table 3.Threshold value for foreground

Table 4.Threshold value for background

The features are grouped both as individual and combined. Individual features show the amount of contrast present in a bitplane and the combined features shows the amount of contrast present between the FG and BG of image. The contrast changes when different sets of combined features from the FG and BG are combined. When a large amount of contrast is present between FG and BG, then clear segmentation can be performed.

Fig. 4 shows the intensity values along the lung border for a 3*3 location. For both individual features and combined features, this graph reveals the possible level of intensity values present at a pixel both in the foreground and background. The intensity values present in the foreground is shown using Diamond and feature values present in the background is shown using Square. The plot shows that a significant slope with a difference of 60 degree in intensity values. Hence, segmentation value can be assigned at a grayscale of 100. This can be compared with the grayscale value mentioned for lung in Table 1.

The membership table is formed based on the fuzzy rules for the chosen block size of 3×3.through the inference process. The bit values assigned at each cell of the membership table is based on the product of bit values presented for UN and LN variables. The numbers in parenthesis in below table are the bit values assigned based on the fuzzy rules shown in the inference process

The Computed intensity values of foreground and background for the proposed method are shown in Figs. 5(a) and (b).

Fig. 5.Computed intensity values: (a) For foreground of 3×3 pixels; (b) for background of the 3x3 pixels

The above figure shows a computed intensity values for variables V8, V7, V6, V5, V4, V3, V2, and V1 respectively for all 9 pixels. From these graphs, it can be observed that the values “V6+V5” provides an optimum value for segmenting the foreground and “V4+V3+V2+V1” provides an optimum value for segmenting the background

Fig. 6.Results of the proposed system

The output bit values obtained from the procedure Bit_assignment for the foreground and background images are shown in Tables 8 and 9. Table 10 presents the output Obtained for only one bit values of each pixel based on the output of step 1 and 2. Table 11 depicts the final segmentation values at the center of 3×3 window. These values 0 and 1 help to get the final segmented lung region Fig. 9 shows the results obtained by applying the proposed FLBP technique on the test images. From this figure, it is evident that this FLBP technique produces much smoother results than the other schemes that have been used earlier. It can also be observed that almost all the existing techniques do not provide better results in which there are overlapping of intensities in lung parenchyma around chest wall. However, the FLBP technique has shown promising results on the different test cases which are shown in Fig. 9.

Table 8.Bit values for UN

Table 9.Bit values for LN

Table 10.Obtained Bit values based on output of step 1 and 2

To enhance the efficiency of proposed segmentation algorithm, the performance measures based on statistical metrics have been computed and are given in Tables 12 and 13 for all the chosen lung images. The lung images for this experiment where chosen from a set of 300 images which are available in the LIDC database. The values shown in these table clearly indicates that there is a considerable improvement in the performance with respect to statistical method for the proposed algorithm over the lung image segmentation. Table 12 shows the total number of pixels present in the segmented and ground truth images.

Table 12.Number of pixels in segmented and ground truth images

Table 13.Segmentation accuracy of segmented and unsegmented images

Table 13 presents three region properties obtained by inbuilt command of Matlab software. The properties are solidity, area, and perimeter. These three values are obtained for segmented and unsegmented images. The total of (Solidity+ Area+ Perimeter) of segmented image is divided by the total of (Solidity+Area+Perimeter) of unsegmented object. The segmentation accuracy calculated by Eq. (7) is as follows:

We proposed method 2 as an alternative to evaluate segmentation more meaningfully when compared to method-1. Ground truth image has been formed manually for all the four images using adobe photoshop. The percentage of segmentation is relative with respect to the ground truth image. The overall segmentation accuracy of 97.52 is obtained from the average value computed from the table.

4.3 Measurement and analysis of over/under segmentation

To evaluate the segmentation performance, we used two error metrics which are used to measure the oversegmentation and under-segmentation rates. We define over segmentation of the region of interest as the region volume that is regarded as tissue in this proposed segmentation method, and we also define the undersegmentation of the region of interest as the tissue that is regarded as region volume.

The over-segmentation and under-segmentation ratios for the 10 datasets are shown in Fig. 7. The average over- segmentation volume error per case is 0.34% of the entire volume, while the average under-segmentation is 1.41%.

Fig. 7.Over segmentation/ under segmentation

4.4 Comparison of segmentation accuracy

Table 14 shows the detailed comparison of segmentation accuracy of the proposed method with two existing CT lung image segmentation techniques proposed by Leader et al [20], Soleymanpour et al [21]. From the table, it can be observed that the segmentation accuracy of the proposed methodology is higher when compared to the other existing methods for segmenting lung region form the Computed tomography lung images.

Table 14.Segmentation accuracy for different techniques

 

5. Conclusions and Future work

In this paper, a new fuzzy logic based bitplane thresholding method has been proposed for segmenting the medical images effectively. This method overcomes the limitations of existing methods where entire gray scale value of a pixel is considered. Fuzzy logic is used for each bit of a gray scale. Hence, more clarity is obtained in deciding the pixel which requires segmentation. To measure the performance of the proposed method, new metrics are proposed which measures the segmentation accuracy based on the number of pixels, properties and over/under segmentation ratios. The major contributions of this paper are improvement in segmentation accuracy and a new metrics for the effective measurement of segmentation accuracy. Future works in this direction can be the proposal of a new technique to analyze the presence of nodules in the segmented medical images