| | Automatic segmentation and recognition of lungs and lesion from CT scans of thoraxReceived 12 February 2008; received in revised form 3 October 2008; accepted 30 October 2008. Abstract In this study, a fully automated texture-based segmentation and recognition system for lesion and lungs from CT of thorax is presented. For the segmentation part, we have extracted texture features by Gabor filtering the images, and, then combined these features to segment the target volume by using Fuzzy C Means (FCM) clustering. Since clustering is sensitive to initialization of cluster prototypes, optimal initialization of the cluster prototypes was done by using a Genetic Algorithm. For the recognition stage, we have used cortex like mechanism for extracting statistical features in addition to shape-based features. The segmented regions showed a high degree of imbalance between positive and negative samples, so we employed over and under sampling for balancing the data. Finally, the balanced and normalized data was subjected to Support Vector Machine (SimpleSVM) for training and testing. Results reveal an accuracy of delineation to be 94.06%, 94.32% and 89.04% for left lung, right lung and lesion, respectively. Average sensitivity of the SVM classifier was seen to be 89.48%. 1. Introduction  The success of radiotherapy (RT) depends upon good target delineation, dose coverage and at the same time low doses to the normal tissues surrounding the target. Clinical outcome of radiotherapy can potentially be improved by increasing the precision of tumour localization and dose delivery during the treatment. The size of the necessary margins, and hence the risk for complications, can be reduced if the set-up uncertainties and internal organ motion, such as those caused by breathing motion, are minimised. Thus we need to know the shape and location of the organs at risk (OAR) and the tumour precisely at the time of treatment planning as well as during treatment, in order to achieve both adequate local tumour control and avoid late adverse effects. Respiratory motion is difficult to compensate for in conventional radiotherapy systems. Therefore, an accurate tracking method to monitor the motion of the tumour is of considerable clinical relevance. Studies by [1], [2] have shown that respiratory movement is rather complex and the relationship between the lung and skin surface motion is affected by many anatomic and physiologic factors. This implies that reduction of internal target margins and efficacy of the free-breathing gating technique is to be assessed for individual cases. Image-guided RT can compensate for organ movement and set-up uncertainties, and improved imaging technology provides us with the possibility of collecting CT data – either by kV- or MV-imaging – with the patient at the treatment coach upfront dose delivery. In tomotherapy, CT data is acquired immediately upfront dose delivery and adjustment of patient position and of beams are feasible before irradiation. Keeping this in mind, we see urgent need for fully automated segmentation method in order to characterize lesion motion and deformation in individual cases for 4D treatment planning and dose delivery. In medical imaging, many methods have been developed and tested on a wide range of applications. Despite these efforts, it remains a daunting task for the system designer to decide which approach is suited for a particular segmentation task. In addition, uncertainties are widely present in the image data because of the noise in acquisition and of the partial volume effects originating from the low resolution of sensors. In particular, borders between tissues are not exactly defined and memberships in the boundary regions are intrinsically fuzzy. Clustering for segmenting human organs and tissues using shape or grey level information is especially challenging due to the changing shape of the organs in a stack of slices and the grey level intensity overlap in soft tissues. On the other hand, normal tissues and organs are expected to have a consistent texture within tissues across multiple slices. In this study, we have used texture information from CT volume scans of lung cancer patients for unsupervised segmentation of lesion and organs at risk. We have concentrated on texture analysis method and in particular, Gabor filtering. Our choice is based on the evidence from psychophysical research [3] indicating that the human brain performs a frequency analysis of images. Our choice of using fuzzy clustering algorithm, in particular, Fuzzy C Means (FCM) has been due to the fact that boundaries of lesion are generally not well defined and seem to be fuzzy. However, in order to fully automate the segmentation process, we need to determine an optimal number of clusters to be used in the clustering algorithm. We have experimented with a number of clusters combinations and decided 15 clusters to be the optimal, empirically. In addition, a Genetic Algorithm is used to optimally initialize the cluster centres. Our approach is particularly appealing and differs from others as it is global in nature, fully automated and independent of user intervention based only upon textural information extracted from image data. As the segmentation is fully automatic, we are not required to supply any grid over slices or marked pathologies beforehand for training or user-based seed initialization. The system looks for naturally existing texture patterns in a varying environment and optimally initializes the cluster centres. Pattern recognition systems usually consider a feature space onto which the observation vector is mapped. The feature vector is then used to decide the class to which the observation vector belongs. Human capability of discriminating objects in a scene is independent of scale, rotation and translation. However, when a machine is supposed to do the same, it has to solve the problem of segmenting the image into uniformly textured regions, locating the borders between different textures and recognizing similar textures even if they are differently shaded. Texture-based recognition process involves two phases: the learning phase and the recognition phase. In the learning phase, the target is to train algorithm used for the texture content of the object present in the training data, which generally comprises of images with known class labels. These features can be scalar numbers, discrete histograms or empirical distributions, and they characterize given textural properties of the images, such as spatial structure, contrast, roughness, orientation, etc. In the recognition phase the texture content of the unknown sample is first described with the same texture analysis method. Then the textural features of the sample are compared to those of the training images with a classification algorithm, and the sample is assigned to the category with the best match. Optionally, the unknown sample can be rejected, if the best match is not sufficiently good according to some predefined criteria. In our approach, we have extracted different features from segmented regions (or patches) including features based upon cortex like mechanism very similar to the one used in the human cortex to process visual information [4]. Class imbalance problems were corrected by using over and under sampling strategies. We identified 12 different classes (as delineated by FCM clustering) and used feature vectors from classes in a multiclass Support Vector Machine (SimpleSVM) for recognizing different regions pertaining to lesion and lungs. Our motivation behind performing automated unsupervised segmentation and recognition of CT images has been to facilitate automatic tracking of the lesion and organs at risk in 4DCT image data set for marker-less breathing synchronised radiation therapy applications. Due to the huge amount of data generated during 4D radiation therapy and the marked inter-observer variability in delineation of lesion, automated segmentation and recognition based upon textural information in the CT images may represent a robust method for automatic delineation and tracking of lesion and organs at risk. 1.1. Previous studies Previous studies on automated texture-based interpretation of chest CT have largely concentrated on early detection of lung cancer [5] including nodule detection [6] and [7], classification of nodules into benign or malignant [8], and measurement of nodule size and growth rate in follow up studies [9]. Other researchers have focussed on the detection and quantification of specific diseases or patterns, such as emphysema [10] or ground glass opacities [11]. Uppaluri et al. [12] have developed a general system for regional classification by using small areas that were classified into one of the six categories based upon 15 statistical and fractal texture features. Delorme et al. [13] have used a pixel-based approach to classify lung tissue in six classes using local texture measures. Uchiyama et al. [14] have presented a system employing neural networks that performed classification into six textural classes. Shyu et al. [15] have developed a system that retrieves reference cases similar to the case at hand from a proven database. In their approach they have combined global and anatomical knowledge, combining features from several pathological regions and anatomical indicators per slice. The regions are however manually delineated rather then automatically detected. In all the studies mentioned above, some sort of grid over slices/ROI marking or marked pathologies beforehand are needed for training, thus a supervised approach is used. Many segmentation algorithms especially in the last decade are proposed with one superior to another in some or the other way, however, none seems to be generally optimal [16]. Due to severe drawbacks of supervised segmentation methods, in clinical practice, an unsupervised approach has been followed in recent years [17]. Many other researchers from different disciplines have turned their attention to segmentation of medical images, and, not surprisingly have embraced a broad spectrum of approaches. Amongst these are boundary-based techniques using active contours, snakes [18] and/or Fourier parameterized models [19]. These employ deformable models to conform to boundaries around the target, typically using energy terms constructed from image gradients, derivatives and balloon forces [20]. Their energy functions are minimised with the risk that the local minima may force unsatisfactory results. Consequently, the model needs to begin near the solution or needs to be supervised. 1.2. Method We begin with pre-processing the image in order to remove noise and enhance contrast to make it suitable for further processing. Median filter was used for noise removal. For contrast enhancement, intensity values in the image were mapped such that 1% of data is saturated at low and high intensities. Next, texture features are extracted from pre-processed image by subjecting the image to Gabor filtering. We used four different frequencies and 5 angles totalling 20 feature images. As the dimension of the feature vector is too large, we used a feature selection procedure. Thus only those feature images were used that gave large contribution. For choosing optimal number of clusters, we have experimented with different indexes [21] and minimising the objective function of the clustering algorithm. However, none seemed to be optimal, generally. So we decided the number of clusters empirically to be 15, which seemed to decipher details and delineate boundaries of lungs and lesions satisfactorily. Additionally, time taken to converge was also reasonable. Next, cluster centres were optimally initialized by using a Genetic Algorithm. Segmented images were labelled with different shade/colour per cluster. The segmented patches from lung and lesion were considerably less than from other areas, so we used data over and under sampling in order to balance the dataset. Finally, for the recognition part, we extracted features from delineated patches and subjected them to SimpleSVM classifier. A flow diagram of the method is described in Fig. 1. 1.2.1. Pre-processing and feature extraction Image was pre-processed in order to enhance it for segmentation. We used a median filter to remove unwanted noise and enhance contrast to prepare it for feature extraction. Next, Gabor filtering was done in order to extract texture features. Conventional texture feature extraction algorithms are grouped in four different methods [22]: statistical, geometrical, model-based and signal processing methods. We followed here the signal processing approach and in particular Gabor filtering. Our basis for such a choice was based upon the evidence from psychophysical research [3] indicating that the human brain does a frequency analysis of images. 1.2.2. Gabor filters The basic even-symmetric Gabor filter oriented at 0° is a band-pass filter with unit response [18] as where u0 is the frequency of a sinusoidal plane wave along the x-axis (i.e. the 0° orientation), and σx and σy are the space constants of the Gaussian envelope along x and y coordinates, respectively. Other orientations are obtained by rotating the reference coordinate system ( x,  y). Frequency and orientation selective properties of Gabor filter are more explicit in its frequency domain representation: where σu= 1/2 πσx, σv= 1/2 πσy and A= 2 πσxσy. The frequency domain representation specifies the amount by which the filter modifies or modulates each frequency component of the output image. The feature images are computed first by submitting each Gabor filtered image to a non-linear transformation: where α= 0.25 corresponds to the sigmoid function as in neural networks and a Gaussian averaging filter is used for removing the noise. Family of Gabor functions have been used for many early vision tasks and they represent many advantages for texture segmentation [23]. Gabor filters have many advantages including: tuneable bandwidths, can be defined to operate over a range of spatial frequency channels, and obey the uncertainty principle in two dimensions [24]. An image can be represented as the sum of output of a bank of filters of 2D Gabor-images. Jain and Farrokhnia [25] have suggested a bank of Gabor filters, i.e. Gaussian shaped band-pass filters, with dyadic coverage of the radial spatial frequency and multiple orientations. This choice of parameters was based upon the relations to models of early vision of mammalians as well as the filters joint optimum resolution in time and frequency [26]. In this study, we have used five radial frequencies  and four orientations 0°, 30°, 60° and 90°. In addition value of σx=2σy=1.12λ is chosen, as in [26], where 1/λ is the preferred spatial frequency. This amounted to a total of 20 feature images. 1.2.4. Fuzzy C Means clustering As an improvement to hard means clustering, Bezdek et al. [28] proposed the objective function where Ω= { xk| k∈[1, n]} be a training set containing n unlabeled samples; Y= { yj|j∈[1, c]} is the set of centres of the clusters; Ej( xk) is a dissimilarity measure (distance measure) between the sample xk and the centre yj of a specific cluster j. U= [ μjk] is the c× n fuzzy c-partition matrix, containing the membership values of all the samples in all the clusters; m∈(1, ∞) is the degree of fuzziness. Thus the clustering problem can be defined as the minimisation of Jm w.r.t. Y under the probabilistic constraint: The FCM algorithm can thus be summarized as iteration of the following formulas: and where, in case of Euclidean space Ej( xk) is defined as Fuzzy clustering methods allow the pixel to belong to several clusters simultaneously, with different degrees of membership. The measure of dissimilarity in FCM is given by the squared distance between each data point and the cluster centre, i.e. the Euclidean distance between them and the distance is weighted by the power of the membership degree at that data point. While using FCM clustering, the parameter values for the fuzziness parameter m, is taken to be 2 and the termination tolerance (ɛ) is set to 0.0001 [29]. Optimal cluster centre initialization is done by using the GA approach as described next. 1.2.5. Genetic Algorithm (GA)-based initialization Genetic Algorithms belong to a class of search techniques that mimic the principles of natural selection to develop solutions of large optimization problems [30]. GA's operate by maintaining and manipulating a population of potential solutions called chromosomes. Each chromosome has an associated fitness value which is a qualitative measure of the goodness of the solution encoded in it. This fitness value is used to guide the stochastic selection of chromosomes which are then used to generate new candidate solutions through crossover and mutation. Crossover generates new chromosomes by combining selections of two or more selected parents. Mutation acts by randomly selecting genes which are then altered; thereby preventing suboptimal solutions from persisting and increases diversity in the population. The process of selection, crossover and mutation continues for a fixed number of generations or until a termination condition is satisfied. In this study, the objective function was reformulated [31], [32] in order to remove dependency on the partition matrix U for FCM clustering. Hathaway and Bezdek [32] have shown that local (V) minimisers of Rm and (U) produce local minimisers of Jm and conversely, the (V) part of the local minimisers of Jm yields local minimisers of Rm. For initializing the cluster centres, we used binary grey code representation as this encoding yields faster convergence in some cases and improves performance over a straightforward binary encoding [33]. The number of generations was selected to be 300 and a 20-bit precision per variable was used. We employed a generation gap of 0.9 and fitness-based reinsertion to implement an elitist strategy. Fitness of each of the individual calculated by using objective function and then ranking is used for fitness minimisation [34]. Selection method for breeding [35] was Stochastic Universal Sampling (SUS). We have used double point crossover [31] with a crossover rate of 90%. Each feature of cluster centre was depicted by a real number and in order to perform crossover, the feature value was converted to integer capturing three decimal places of precision and then to binary string. Mutation rate was calculated as in [36] by where P is the population size and bits is the encoding bits per individual population member. In our case we have chosen maximum population size equal to the number of clusters (15). Our stopping criterion for GA was number of generations (300). After determining the optimal seeds for initialization, these were utilized for initializing FCM cluster centres. 1.2.7. Class imbalance and feature normalization Before actually performing any kind of training and classification based upon the features extracted, we needed to determine what kind of data we are working on. From the regions segmented, we found a clear overrepresentation of regions other than lung and lesion. This overrepresentation made the classifier to predict the most common class. Thus the different classes were having imbalanced distribution of samples, i.e. one or more of the classes outnumbered the others in amount of samples presented to the classifier for training. In such cases, most of the classifiers learn the majority sample class and thus lead to poor classification accuracy. One of the common and effective methods to deal with the imbalance is sampling [39]. There are two basic sampling methods: (1)Under sampling of the majority class samples: where majority class samples are reduced. We reduced the majority class samples and balanced the dataset as described below:(a)Exclusion of unwanted patches: here we filtered the regions to be used for training to be the ones which have their centroid falling inside the convex hull [37] of the left and right lung. We assumed that all regions pertaining to lesion and lungs will have their centroid located inside the convex hull of the left and right lung. (b)Clinically insignificant tumour: in a further pruning step, we have filtered away patches which were having less than 5 pixels per patch, as we assume that clinically significant tumour should be at least 5 pixel big. (c)Inclusion of similar patches as to lung and lesion: lung and Lesion normally represent a circular to a semi-circular horseshoe shaped patch. Lesion is seen normally as a more circular patch whereas lungs more of an elliptical. Considering this and experimenting with different values of shape factor (see Eq. (13)), we found that a lower threshold for shape as 0.15 could be used safely, for further reducing in data:By using the above-mentioned criteria, we reduced the majority samples in order to balance with the minority samples. This was motivated by [40] where the author indicates that optimal learning is achieved via a balanced learning set. (2)Over sampling of the minority class samples: motivation for using over sampling comes from the observation that natural distribution of the samples is not likely to have optimal distribution of samples for each and every class as seen from a classification perspective. Previous study [41], has also indicated that accuracy degradation on unbalanced set is more severe as compared to case where there is overlap seen between the features. We have utilized here the SMOTE technique for over sampling [42] all the balanced dataset. (3)Feature normalization: in classification tasks, features that characterize the samples quite often have different ranges. Many classifiers including SVM and neural networks require that normalization of the features is done so that their values fall between a certain ranges. In SVM, it is important pre-processing step in order to prevent features with large ranges from dominating the calculations. The most common normalization method is the z-score transformation [43] which guarantees 99% of features to be in the range [0,1]. Thus z-score transformation was used for normalization of the features in range [0,1]. 2. SimpleSVM classifier This subsection presents the theory of SimpleSVM based upon [44]. The binary discrimination SVM problem with a sample (xi,yi) for i=1, …, m with labels y∈{−1,1} is given as the solution of optimization problem under constraints as where i= 1, …, m and ξi≥ 0 is a slack variable. C is a scalar that adjusts the smoothness of the decision function and b is scalar called as bias. The solution to this problem is given by the saddle point of the Lagrangian: from which we retrieve part of Kuhn–Tucker conditions: By using  in (15) we can eliminate f in the Langrangian and end up with dual formulation: with constraints αty= 0 and 0 ≤ αt≤ C. G is the influence Matrix of general term Gij= yiyjk( xi, xj) and ɛ= [1, …, 1] T. k( xi, xj) is the kernel function and specifies the non-linear mapping of the data. In this study, we have used the RBF Kernel [45] as linear Kernel cannot handle non-linear separable classification task and computationally stable than other kernels like polynomial kernel and Sigmoid Kernels. The RBF Kernel is given by where γ is the hyperparameter to be determined. In our simulations, we found that γ= 0.9 gave optimal results. For multiclass classification, we used the ‘1vsall’ formulation, where SVM is trained for each class considering all other classes as one. The final decision was given by taking the maximum of the classifiers. 3. Materials The CT-scans were acquired from patients with lung cancer at the Rikshospitalet Radiumhospitalet Medical Centre, Oslo, Norway. The CT-scan images were acquired on a 4-slice CT scanner (GE) and a single slice CT scanner (GE) equipped with fluoroscopy functionality. Some of the images are acquired during fluoroscopy guided biopsy, while others are acquired as part of a diagnostic examination and some are acquired during radio frequency ablation (RFA). For images acquired from diagnostic examination, we have used a LightSpeed QX/i4 slice CT scanner with pitch of 1.5, slice thickness of 2.5mm, in helical mode. For fluoroscopy guided biopsy and RFA, HiSpeed CT/i was used with slice thickness 3mm, pitch 0 and helical mode. Images are stored in Dicom (Digital Communication and Imaging in Medicine) format, of size 512×512 and with 12-bit grey level resolution. The Images were anonymized and exported from PACS (Agfa). Another set of image used was obtained from The Finsens Centre, Rigshospital, Copenhagen. The CT-scan was acquired from a patient with lung cancer, undergoing radiation treatment planning. The patient underwent a 4D CT-scan as a part of a clinical study. The CT-scan images were acquired on a 4-slice CT scanner (GE PET/CT scanner Discovery LS) equipped with the Real-Time Position Management (RPM™) Respiratory Gating system (Varian Medical Systems, Inc.) The patient was audio-coached into at regular breathing pattern as close to normal breathing as possible. The scan mode was sequential with a rotation time of 0.5s, obtaining images in more than one full respiratory cycle in each bed position and after completion of the scan; the images were retrospectively sorted in to 10 bins, according to the breathing signal recorded by the RPM™ system. This operation was performed with the advantage 4D software from GE Medical Systems. Each bin represents 1/10 of the full respiratory cycle, for this patient 5.5s on average for a full cycle (range 4.6–6.5s) equalling 0.56s in each bin. Most of the images used for analysis (41) were obtained from the Rikshospital-Radiumhospitalet medical Centre, Norway all from different patients. Only one image was used from 4D CT Dataset from Finsen Centre, Rigshospital, Copenhagen. Simulations were carried on a 2.53GHz Pentium Processor® with a MS Windows XP® operative system with the help of Matlab® version 2006b. Images used were in Dicom format, of size 512×512 and with 12-bit grey level resolution. Software is developed in Matlab®. 4. Results and discussion In order to prepare the image for segmentation, pre-processing of the image was done by contrast enhancement and median filtering as shown in Fig. 2. Median filter was used for noise removal. Contrast enhancement was performed by mapping the intensity values to new values such that 1% of data is saturated at low and high intensities. Next, a total of 20 Gabor filtered feature images along with grey values and XY coordinate were extracted and used to form the feature vector. These images were subjected to a feature selection method which resulted in the six most contributing feature images (as in this case) shown in Fig. 3. These feature images were combined along with XY coordinate information and pixel value in order to form a feature vector which was then normalized. As sufficient number of clusters is needed prior to clustering, we choose the number of clusters to be 15, empirically. As FCM is also sensitive to initialization, we used Genetic Algorithm to determine optimal cluster centres (prototypes). Combining GA and FCM, we clustered 42 images from different patients with varied tumour size and location. The feature vector was normalized, before clustering, by using the z-score transformation in the range [01]. Fig. 4 shows pre-processed images and Fig. 5 shows corresponding segmented images. We identified 12 different classes/categories of segmented patches (see Table 2). The individual classes representing left lung (LLUP, LLIP, LLLO, LLOP), right lung (RLUP, RLIP, RLLO, RLOP) and lesion (LEIN1, LEIP1 and LEOP1) are then merged (added) for each region, respectively. Finally any holes falling inside the region boundaries were also removed by running a fill hole algorithm [38]. The merged regions were then compared with the expert radiologist delineations (see Fig. 6). The ROC plots are shown in Fig. 7. The accuracy of delineation was seen to be 94.06%, 94.32% and 89.04% for left lung, right lung and lesion, respectively. The area under the ROC curve (AUC) is seen to be 0.9746, 0.9937 and 0.9548 for left lung, right lung and lesion, respectively. | | |  | # | CLASS | Meaning | |
|---|
| 1 | LLUP | Left Lung Upper Part or Inner Part | | | 2 | LLLO | Left Lung Lower Part | | | 3 | LLIP | Left Lung Inner Periphery | | | 4 | LLOP | Left Lung Outer Periphery | | | 5 | RLUP | Right Lung Upper Part or Inner Part | | | 6 | RLLO | Right Lung Lower Part | | | 7 | RLIP | Right Lung Inner Periphery | | | 8 | RLOP | Right Lung Outer Periphery | | | 9 | LEIN1 | Lesion Innermost core | | | 10 | LEOP1 | Lesion Outer Periphery | | | 11 | LEIP1 | Lesion Inner Periphery | | | 12 | BACK | Background Regions | | | | |
Next, we have used a Cortex like mechanism (see Fig. 8) along with shape and position-based features (see Table 1) in order to form a feature vector, for automatic recognition of delineated regions from texture-based segmentation. It was observed that amount of feature vectors were unbalanced and patches delineated from regions other than lungs and lesion was more in number. Thus, we needed to balance the data set, prior to classification. We used Convex Hull, area and shape factor to undersample the feature vectors from classes other than lung and lesion. For over sampling the data set, we used the SMOTE algorithm. After balancing the data set, a training set using 21 segmented images and a test set using the remaining 21 images (all from different patients) were used. By using over and under sampling, 500 samples from each class were generated for training and testing purposes. SimpleSVM algorithm was used as a classifier. Results from classification experiments are presented in Table 3. On an average we achieved a sensitivity of 89.48% for all the classes. | | | | Actual vs. predicted classes | LLUP | LLLO | LLOP | LLIP | RLUP | RLLO | RLOP | RLIP | LEIN1 | LEOP1 | LEIP1 | BACK | |
|---|
| LLUP | 97.40 | – | – | 2.60 | – | – | – | – | – | – | – | – | | | LLLO | – | 94.0 | – | – | 2.80 | 3.20 | – | – | – | – | – | – | | | LLOP | – | – | 98.80 | – | – | – | 1.20 | – | – | – | – | | | | LLIP | – | – | – | 100 | – | – | – | – | – | – | – | – | | | RLUP | – | 7.80 | – | – | 86.60 | 3.20 | – | – | – | – | – | 2.40 | | | RLLO | – | – | – | – | | 100 | – | – | – | – | – | – | | | RLOP | – | – | – | 0.20 | – | – | 5.60 | 94.20 | – | – | – | – | | | RLIP | – | – | – | 0.60 | – | – | 4.0 | 94.60 | 0.80 | – | – | – | | | LEIN1 | – | – | – | – | – | – | 2.20 | – | 97.80 | | | – | | | LEOP1 | – | – | – | – | – | – | – | – | – | 100 | – | – | | | LEIP1 | – | – | – | – | – | – | – | – | 1.0 | – | 99.0 | – | | | BACK | – | – | – | – | – | – | – | – | – | – | – | 100 | | | | |
The main disadvantage of our approach is that it takes substantial amount of processing time for segmentation even on modern computers (approx. 2h). Secondly, the classifier seems to confuse between the outer periphery of the right lung with the inner periphery of the right lung (see Table 3). We are not yet sure if it is a disadvantage; however, we are looking into other algorithms in order to improve the system performance. In our future work, we propose to work on predicting the different classes of lesion and regions of lungs from image frame to frame in a 4DCT dataset based upon automatic segmentation and recognition as described in this work. 5. Conclusions Based upon texture features, as extracted from Gabor filtering, we see that FCM can be used for segmentation of CT of thorax given that the cluster centres are initialized by using a Genetic Algorithm. From the segmentation results, we were able to identify 12 different classes. The accuracy of delineation was seen to be 94.06%, 94.32% and 89.04% for left lung, right lung and lesion, respectively. For automatically recognizing the segmented regions, an average sensitivity of 89.48% was achieved by combining cortex-like, shape and position-based features in a SimpleSVM classifier. Acknowledgements We would like to take this opportunity to thank Radiograph Mats Alexander Olsen for anonymizing the Dicom Images so that they could be used for analysis; Dr. Trond Hagtvedt for providing patient database; Prof. Øyvind Andreassen for helpful discussions and Prof. Øyvind S. Bruland for clinical contributions and finally Ms. Sukhdeep for proofreading the manuscript. References [1]. [1]Liu HH, Koch N, Starkschall G, Jacobson M, Forster K, Liao Z, et al. Evaluation of internal lung motion for respiratory-gated radiotherapy using MRI. Part II. Margin reduction of internal target volume. Int J Radiat Oncol Biol Phys. 2004;60(Dec (5)):1473–1483. Abstract | Full Text |
Full-Text PDF (667 KB)
|
CrossRef
[2]. [2]Koch N, Liu HH, Starkschall G, Jacobson M, Forster K, Liao Z, et al. Evaluation of internal lung motion for respiratory-gated radiotherapy using MRI. Part I. Correlating internal lung motion with skin fiducial motion. Int J Radiat Oncol Biol Phys. 2004;60(Dec (5)):1459–1472. Abstract | Full Text |
Full-Text PDF (701 KB)
|
CrossRef
[3]. [3]Julesz B. Nonlinear and cooperative processes in texture perception. In: Theoretical approaches in neurobiology. Cambridge, MA: MIT press; 1981;p. 93–108. [4]. [4]Serre T, Wolf L, Bileschi S, Riesenhuber M, Poggio T. Robust object recognition with cortex-like mechanisms. IEEE Trans Pattern Anal Mach Intell. 2007;29(Mar (3)):411–426. MEDLINE |
CrossRef
[5]. [5]Reeves AP, Kostis WJ. Computer aided diagnosis for lung cancer. Radiol Clin North Am. 2000;38:497–509. Full Text |
Full-Text PDF (1080 KB)
|
CrossRef
[6]. [6]Armato SG, Giger ML, MacMohan H. Automated detection of lung nodules in CT scans: preliminary results. Med Phys. 2001;28:1552–1561. MEDLINE |
CrossRef
[7]. [7]Lee Y, Hara T, Fujita H, Itoh S, Ishigaki T. Automated detection of pulmonary nodules in helical CT images based on an improved template-matching technique. IEEE Trans Med Imaging. 2001;20:595–604. MEDLINE |
CrossRef
[8]. [8]McNitt-Gray MF, Har EM, Wyckoff N, Sayre JW, Goldin JG. A pattern classification approach to characterizing solitary pulmonary nodules imaged on high resolution CT: preliminary results. Med Phys. 1999;26:880–888. MEDLINE |
CrossRef
[9]. [9]Yankelevitz DF, Reeves AP, Kostis WJ, Zhao B, Henschke CI. Small pulmonary nodules: volumetrically determined growth rates based upon CT evaluation. Radiology. 2000;217:251–256. MEDLINE [10]. [10]Zagers H, Vrooman HA, Aarts NJM, Stolk J, Kool LJS, Dijkman JH, et al. Assessment of the progression of emphysema by quantitative analysis of spirometrically gated computed tomography images. Invest Radiol. 1996;31:761–767. MEDLINE |
CrossRef
[11]. [11]Kauczor HU, Heitmann K, Heussel CP, Marwede D, Uthmann T, Thelen M. Automatic detection and quantification of ground-glass opacities on high resolution CT using multiple neural networks: comparison with a density mask. Am J Roentgenol. 2000;175:1329–1334. [12]. [12]Uppaluri R, Hoffman EA, Sonka M, Hartley PG, Hunninghake GW, McLennan G. Computer recognition of regional lung disease patterns. Am J Respir Crit Care Med. 1999;160:648–654. [13]. [13]Delorme S, Keller-Reichenbecher MA, Zuna I, Schlegel W, van Kaik G. Usual interstitial pneumonia—quantitative assessment of high-resolution computed tomography findings by computer-assisted texture based image analysis. Invest Radiol. 1997;32:566–574. MEDLINE |
CrossRef
[14]. [14]Uchiyama Y, Katsuragawa S, Abe H, Shiraishi J, Li F, Li Q, et al. Quantitative computerized analysis of diffuse lung disease in high-resolution computed tomography. Med Phys. 2003;30:2440–2454. MEDLINE |
CrossRef
[15]. [15]Shyu CR, Brodley CE, Kak AC, Kosaka A, Aisen AM, Broderick LS. ASSERT: a physician-in-the-loop content based retrieval system for HRCT image databases. Comput Vis Image Underst. 1999;75:111–132. [16]. [16]Udupa JK, Saha PK. Fuzzy connectedness and image segmentation. Proc IEEE. 2003;91(10):1649–1669. [17]. [17]Bezdek JC, Hall LO, Clarke LP. Review of MR image segmentation techniques using pattern recognition. Med Phys. 1993;20(Jul (4)):1033–1048. MEDLINE |
CrossRef
[18]. [18]Farrokhnia F, Jain AK. A multi-channel filtering approach to texture segmentation. In: Proceedings CVPR ‘91, IEEE conference on computer vision and pattern recognition. Maui, HI, USA. 1991;p. 364–370. [19]. [19]Cohen I, Cohen LD, Ayache NJ. Using deformable surfaces to segment 3-d images and infer differential structures. CVGIP: Image Underst. 1992;56(2):135–263. [20]. [20]Worring M, Smeulders AWM, Staib LH, Duncan JS. Parameterized feasible boundaries in gradient vector fields. Comput Vis Image Underst. 1996;63(Jan (1)):135–144. [21]. [21]Kakar M, Nystrom H, Nøttrup TJ, Bruland OS, Olsen DR. Automated texture based CT segmentation by Gabor filtering and Fuzzy Clustering. Med Phys. 2006;3(6):2170. [22]. [22]Tuceryan M, Jain AK. Texture analysis. In: Chen CH, Pau LF, Wang PSP editor. Handbook of pattern recognition and computer vision. 2nd edn.. World Scientific Publishing Co.; 1998;p. 207–248. [23]. [23]Field DJ. Relations between the statistics of natural images and the response properties of cortical cells. J Opt Soc Am A. 1987;4(12):2379. MEDLINE [24]. [24]Turner MR. Texture segmentation by Gabor functions. Biol Cybern. 1986;55:71–82. MEDLINE [25]. [25]Jain AK, Farrokhnia F. Unsupervised texture segmentation using Gabor filters. In: International conference on systems, man and cybernetics. Los Angeles, CA, USA. 1990;p. 14–19. [26]. [26]De Valois RL, Albrecht DG, Thorell LG. Spatial frequency selectivity of cells in macaque visual cortex. Vision Res. 1982;22(5):545–559. MEDLINE |
CrossRef
[27]. [27]Farrokhnia F, Jain AK. A multi-channel filtering approach to texture segmentation. In: Computer Vision and Pattern Recognition, 1991. Proceedings CVPR ‘91, IEEE Computer Society Conference. 1991;p. 364–370. [28]. [28]Bezdek J, Keller J, Krishnapuram R, Pal NR. Fuzzy models and algorithms for pattern recognition and image processing. Boston: Kluwer Academy Publishers; 1999;. [29]. [29]Pal NR, Bezdek JC. On cluster validity for the fuzzy c-means model. IEEE Trans Fuzzy Syst. 1995;3(3):370–379. [30]. [30]Goldberg DE. Genetic Algorithms in search, optimization and machine learning. New York: Addison-Wesley; 1989;. [31]. [31]Hall LO, Ozyurt IB, Bezdek JC. Clustering with a genetically optimized approach. IEEE Trans Evol Comput. 1999;3(2):103–112. [32]. [32]Hathaway RJ, Bezdek JC. Optimization of clustering criteria by reformulation. IEEE Trans Fuzzy Syst. 1995;3(2):241–245. [33]. [33]Back T, Kursawe F. Evolutionary algorithms for fuzzy logic: a brief overview. In: Proceedings of information processing and management of uncertainity in knowledge-based systems, vol. 2. 1994;p. 659–664. [34]. [34]Whitley D. The GENITOR algorithm and selection pressure: why ranking based allocation of reproductive trials is best. Proc IGGA. 1989;3:116–121. [35]. [35]Baker JE. Reducing bias and inefficiency in the selection algorithm. Proc ICGA. 1987;2:14–21. [36]. [36]Schaffer JD, Caruna RA, Eshelman LJ, Das R. A study of control parameters affecting online performance of Genetic Algorithms for function optimization. In: Proceedings of the third international conference on Genetic Algorithms. 1989;p. 51–60. [37]. [37]Linda G, Shapiro , George C, Stockman . Computer vision. Upper Saddle River, NJ: Prentice Hall; 2001;. [38]. [38]Gozalez RC, Woods RE. Digital image processing. MA: Addison-Wesley Publishing Company; 1993;. [39]. [39]Weiss Gary M. Mining with rarity: a unifying framework. SIGKDD Explor Newsl. 2004;7–9. [40]. [40]Weiss GM, Provost F. Learning when training data are costly: the effect of class distribution on tree induction. J Artif Intell Res. 2003;19(Oct):315–354. [41]. [41]Prati RC, Batista GEAPA, Monard MC. Class imbalances versus class overlapping: an analysis of a learning system behavior. Lect Notes Comp Sci. 2004;2972:312–321. [42]. [42]Chawla NV, Boywer KW, Hall LO, Kegelmeyer WP. SMOTE: synthetic minority over-sampling technique. J Artif Intell Res. 2002;(Jun (16)):321. [43]. [43]Aksoy S, Haralick RM. Feature normalization and likelihood-based similarity measures for image retrieval. Pattern Recogn Lett. 2001;22(Apr (5)):563–582. [44]. [44]Vishawanathan SVM, Smola AJ, Murty MN. In: Fawcett T, Mishra N editor. SimpleSVM. Washington DC, USA: AAAI Press; 2003;p. 760–767. [45]. [45]Burges CJC. A tutorial on support vector machines for pattern recognition. Data Min Knowl Discov. 1998;2(2):121–167. [46]. [46]Udupa JK, LeBlanc VR, Zhuge Y, Imielinska C, Schmidt H, Currie LM, et al. A framework for evaluating image segmentation algorithms. Comput Med Imaging Graph. 2006;30(Mar (2)):75–87. Abstract | Full Text |
Full-Text PDF (467 KB)
|
CrossRef
[47]. [47]Huang J, Ling CX. Using AUC and accuracy in evaluating learning algorithms. IEEE Trans Knowl Data Eng. 2005;17(Mar (3)):299–310. [48]. [48]Kohavi R, Provost F. Applications of data mining to electronic commerce. Data Min Knowl Discov. 2001;5(1–2):5–10. Mr. Manish Kakar received his BSc(H) degree in Physics from University of Delhi, India in 1993. He received his MSc in Electrical Engineering from University of Tromsoe, Norway in 1998. He is currently pursuing his PhD degree at the University of Oslo, Norway. His current research includes biomedical signal and image processing, pattern recognition, soft computing and Image Guided Radiation Therapy (IGRT). Dr. Dag Rune Olsen received his PhD degree from University of Oslo, Norway. He is an associate professor at the Department of Physics, University of Oslo in addition to being Head of the Department of Radiation Biology, Rikshospitalet-Radiumhospitalet Medical Centre. His research interests include BioART (Biologically Adaptive Radiation Therapy) and Medical Imaging. a Department of Radiation Biology, Institute for Cancer Research, Rikshospitalet-Radiumhospitalet Medical centre, Oslo, Norway b Department of Medical Physics, Institute for Cancer Research, Rikshospitalet-Radiumhospitalet Medical centre, Oslo, Norway c Department of Physics, University of Oslo, Oslo, Norway  Corresponding author. Tel.: +47 22781225; fax: +47 22781207.
PII: S0895-6111(08)00118-3 doi:10.1016/j.compmedimag.2008.10.009 © 2008 Elsevier Ltd. All rights reserved. |
 11 of 11
|
| |
|
FULL TEXT00118-3/journalimage?img=PIIS0895611108001183.gr1.lrg.gif&fig=fig1&kwhquery=&issn=0895-6111&ishighres=false&allhighres=false&free=yes','journalimage');)
00118-3/journalimage?img=PIIS0895611108001183.gr2.lrg.gif&fig=fig2&kwhquery=&issn=0895-6111&ishighres=false&allhighres=false&free=yes','journalimage');)
00118-3/journalimage?img=PIIS0895611108001183.gr3.lrg.gif&fig=fig3&kwhquery=&issn=0895-6111&ishighres=false&allhighres=false&free=yes','journalimage');)
00118-3/journalimage?img=PIIS0895611108001183.gr4.lrg.gif&fig=fig4&kwhquery=&issn=0895-6111&ishighres=false&allhighres=false&free=yes','journalimage');)
00118-3/journalimage?img=PIIS0895611108001183.gr5.lrg.gif&fig=fig5&kwhquery=&issn=0895-6111&ishighres=false&allhighres=false&free=yes','journalimage');)
00118-3/journalimage?img=PIIS0895611108001183.gr6.lrg.gif&fig=fig6&kwhquery=&issn=0895-6111&ishighres=false&allhighres=false&free=yes','journalimage');)
00118-3/journalimage?img=PIIS0895611108001183.gr7.lrg.gif&fig=fig7&kwhquery=&issn=0895-6111&ishighres=false&allhighres=false&free=yes','journalimage');)
00118-3/journalimage?img=PIIS0895611108001183.gr8.lrg.gif&fig=fig8&kwhquery=&issn=0895-6111&ishighres=false&allhighres=false&free=yes','journalimage');)