Automated detection of lung nodules in computed ... - Springer Link

Machine Vision and Applications (2012) 23:151–163 DOI 10.1007/s00138-010-0271-2

ORIGINAL PAPER

Automated detection of lung nodules in computed tomography images: a review S. L. A. Lee · A. Z. Kouzani · E. J. Hu

Received: 4 December 2009 / Revised: 19 April 2010 / Accepted: 22 April 2010 / Published online: 14 May 2010 © Springer-Verlag 2010

Abstract Lung nodules refer to a range of lung abnormalities the detection of which can facilitate early treatment for lung patients. Lung nodules can be detected by radiologists through examining lung images. Automated detection systems that locate nodules of various sizes within lung images can assist radiologists in their decision making. This paper presents a study of the existing methods on automated lung nodule detection. It introduces a generic structure for lung nodule detection that can be used to represent and describe the existing methods. The structure consists of a number of components including: acquisition, pre-processing, lung segmentation, nodule detection, and false positives reduction. The paper describes the algorithms used to realise each component in different systems. It also provides a comparison of the performance of the existing approaches. Keywords Computed tomography · Lung images · Pulmonary nodules · Automated detection · Performance evaluation 1 Introduction Lung cancer is caused by uncontrollable irregular growth of cells in lung tissue. Early detection of lung tissue abnormalS. L. A. Lee · A. Z. Kouzani (B) School of Engineering, Deakin University, Geelong, VIC 3217, Australia e-mail: [email protected] S. L. A. Lee e-mail: [email protected] E. J. Hu School of Mechanical Engineering, Adelaide University, North Terrace, Adelaide, SA 5005, Australia e-mail: [email protected]

ities can help lung patients getting early treatment for their illness. Lung tissue abnormalities that are roughly spherical with round opacity and a diameter of up to approximately 30 mm [1] are known as lung nodules. They can be categorised into a number of groups [2,3] including: juxta-vascular, well-circumscribed, pleural tail, and juxta-pleural. Juxta-vascular has significant connections to its neighbouring vessels. Well-circumscribed has no connection to its neighbouring vessels and structures. Pleural tail has thin connection to its neighbouring pleural wall, where the pleural tail belongs to the nodule itself. Juxta-pleural has some degree of attachment to its neighbouring pleural surface. Figure 1 shows a sample image for each of the stated nodule categories. Advancements in computed tomography (CT) have offered a considerable opportunity in helping lung cancer patients getting early treatment for their illness. This is because CT enables visualisation of small or low-contrast nodules that could hardly be screened in conventional radiograms. The study presented in this paper aims to identify and describe the existing methods on automated and semi-automated lung nodule detection. It presents a review of the published work on nodule detection. Whilst there exists a number of good review papers on lung nodule detection including [4–9], these papers focus on a subset of the available literature. While [4] gives a very extensive review of the existing works on nodule detection, it only covers the published works up to 2004. The current paper, however, provides a comprehensive review of the existing work to date covering the reported modern lung nodule detection methods. An investigation of the existing publications on lung nodule detection reveals that a proper attempt to categorise the existing nodule detection methods based on their operation principles has not been made. This paper, however, devises a generic structure for lung nodule detection that can be used to categorise and describe majority of the existing methods.

123

152

S. L. A. Lee et al.

Fig. 1 Sample images for the four nodule groups. From left to right juxta-vascular, well-circumscribed, pleural tail, and juxta-pleural nodules

It consists of a number of components including: acquisition, pre-processing, lung segmentation, nodule detection, and false positives reduction. Various algorithms have been employed to realise each component in different systems. These algorithms are reviewed in this paper. In addition, the paper provides a comparison of the performance of the existing approaches making it easier for the reader to establish an understanding of the applicability of the studied approaches.

2 Review of existing nodule detection methods There are many articles that have described methods for automated and semi-automated detection of pulmonary nodules. A study of the developed methods reveals that the methods employ different structures. Each structure involves a number of algorithmic components as well as their specific inter-relationships. This paper devises a generic structure for lung nodule detection that can be used to categorise and represent majority of the existing approaches. This structure is displayed in Fig. 2. It consists of a number of components including: acquisition, pre-processing, lung segmentation, nodule detection, and false positives reduction. Some

Fig. 3 A sample lung image from LIDC

existing nodule detection systems include all these components, whilst others employ only a subset of the components. When a system does not include a certain component, the presented generic structure could be reduced by bypassing the particular component and creating a smaller structure for the system. Presentation of the existing systems based on the devised structure helps the reader to better establish an understanding of the operation principles of the systems, and also compare the characteristics of the methods that employ similar frameworks. In the following, each component is first defined, and then the existing methods relating to each component are described. 2.1 Acquisition Image acquisition refers to the process of acquiring medical images from imaging modalities. There exist several common methods for lung imaging. Computed tomography (CT) enables visualisation of small volume or low-contrast nodules by decreasing the thickness of slices and the interval between consecutive slices. CT is preferable for the preliminary analysis of lung nodules screening comparing to other lung imaging methods. Lung CT images can be found in public and private databases. There are a number of popular public lung nodule databases including: Early Lung Cancer Action Program (ELCAP) Public Lung Image Database [10], ELCAP Public Lung Database to Address Drug Response [11], Lung Image Database Consortium (LIDC) in National Imaging Archive [12] (see Fig. 3), and Medical Image Database [13]. Whilst many reported lung nodule detection methods [2,3,14–19] employed CT images from these public databases, according to the literature, numerous researchers used private databases obtained from their partner hospitals [6–9]. 2.2 Pre-processing

Fig. 2 The generic structure of lung nodule detection methods

123

Image pre-processing refers to the process of improving both the quality and interpretability of the acquired lung images.

Automated detection of lung nodules in CT images

Fig. 4 A sample pre-processed lung image: (left) original and (right) pre-processed images

The pre-processing component reduces noise and artefacts in the lung image slices (see Fig. 4). A low-pass filter with disk and Gaussian parameters was explored by Garnavi et al. [20]. Smoothing was implemented by Kim et al. [21] to reduce noise through median filtering. This method was also used in the work of Gurcan et al. [22], Lin and Yan [23], and Lin et al. [24]. Gaussian smoothing was employed by Pu et al. [25], Wei et al. [26], Gori et al. [27], and Retico et al. [28] to eliminate the image artefacts. This removes the small contour along the lung boundary which is falsely regarded as the lung boundary. Kawata et al. [29] utilised a Gaussian smoothing filter to enhance the image slices before applying a deformable model. Arimura et al. [30] proposed a ring average filter to suppress nodule image, and a matched filter to enhance the nodule image. A linear feature detector filter and a Laplacian filter were employed to remove streak shadows by Kubo et al. [31]. A Laplacian of Gaussian (LoG) filter was proposed by Diciotti et al. [3] and Sluimer et al. [32] for enhancement of lung images. Bae et al. [33] used a morphological filter to enhance the image region. In Ochs et al. [15] and Paik et al. [34] studies, a sphericity structure enhancement filter was applied to enhance the nodule like structure in CT images. Oda et al. [35] presented a 3D filter to map the gradient vectors orientation of the ground glass opacity nodules. Conformal nodule filtering and un-sharp masking were described by Farag et al. [36,37] to enhance nodules candidates and suppress other structures. Zhao et al. [38] utilised 3D cylinder filter and Chang et al. [39] combined the cylinder filter with a spherical filter to suppress the vessel and other extended parts including noises within the lung region, while preserving the ground glass opacity intensity values. An anisotropic filter was utilised by Kostis et al. [2] to enhance different classes of nodules. An enhancement method known as quantised convergence index (QCI) filter was presented by Matsumoto et al. [40]. An N-Quoit filter was implemented by Ezoe et al. [41] and Tanino et al. [42] to enhance candidate nodules. Top hat and sieve filters were explored by Awai et al. [43], and then com-

153

Fig. 5 A sample segmented lung image: (left) original and (right) segmented images

bined with a partial reconstruction filter by Fukano et al. [44] to highlight the vessel ridge shadow. A multi-scale enhancement filter was used by Li and Doi [45], Li et al. [46], and Yu et al. [47] to suppress blood vessels and highlight nodule like structures. A selective filter based on enhancing dot-like features (nodule) and suppressing line-like features (vessel) were proposed by Fetita et al. [48], Korfiatis et al. [19], Sun et al. [49], and Jia et al. [50]. 2.3 Lung segmentation Lung Segmentation refers to the process of identifying the lung lobe region and removing the rest of the image (see Fig. 5). It plays a crucial role in pulmonary nodule detection by increasing reliability, accuracy, and precision, and decreasing computational cost of detection. Dynamic programming was used by Xu et al. [51] and Aoyama [52] to isolate nodule regions in images. Also, a twophase method involving dynamic programming and spiral scanning was employed by Wang et al. [14] to outline nodule regions. A deformable model was used by Kawata et al. [29,53] to segment nodules images. A combination of thresholding, region filling, and deformable model was developed by Kim et al. [54] to segment the lung region. Bellotti et al. [55] employed region growing with contour following to isolate juxta-pleural nodules. The deformable model had issues related to speed, initialisation, and poor convergence on border cavities. According to Ref. [54], inclusion of a thresholding with the deformable model could resolve the issues. The snake algorithm which is an active contour model was implemented by Itai et al. [56] to obtain the boundary of nodules. Zhao et al. [57] improved the shape-based segmentation using nodule gradient and sphere occupancy measurements. Ko et al. [58] formed a thresholding and partial volume method for volume quantisation of nodules. Okada et al. [59] proposed a multi-scale Gaussian intensity model fitting for segmentation of ellipsoidal nodules. Fan et al. [60] developed a 3D template which initialises and analyses the nodule cross correlation curve. A morphology-based approach was employed by Kostis et al. [2] and Enquobahrie et al. [61].

123

154

It removes the irrelevant structures that may be present in the image. Morphological opening, erosion, thresholding, seed optimisation, and boundary refinement operations were used by Kuhnigk et al. [62] to extract large nodules. Thresholding and morphology operations were also used by Kanawaza et al. [63] to extract lung regions and reduce partial volume and beam hardening effects. Thresholding, Canny edge detection, and morphological closing were utilised by Paik et al. [34] and Korfiatis et al. [19] to segment juxta-pleural nodules. Region growing was explored by Diciotti et al. [3], Lee et al. [64,65], and Shah and McNitt-Gray et al. [66,67] for lung tissue segmentation. Combining the region growing with morphological operations, Lin and Yan [23] and Lin et al. [24] achieved filling of the indentation caused by blood vessel that could not be handled by thresholding. A fuzzy connectivity method was developed by Udupa et al. [68] which focuses on the strongest affinity path between each specific point and the seed point. Dehmeshki et al. [69] formed a combination of the region growing and fuzzy connectivity methods to segment nodules. Also, Klik et al. [70] combined the region growing with Hough transform to extract the nodule candidates. Thresholding and seeded segmentation were performed by Bae et al. [33] to isolate the juxta-pleural nodule from other structures. Awai et al. [43] performed thresholding and connected-component labelling to segment candidate nodule regions. Simple thresholding was presented in Farag et al. [36,37], El-Baz et al. [71], and Giger et al. [72] for separation of the nodule candidates from the background image. According to Zhao et al. [38], simple thresholding and edge segmentation do not result in an acceptable segmentation for the ground glass opacity. A 3D texture likelihood map using non-parametric density estimation was formed to evaluate the presence of the ground glass opacity voxels. Wang et al. [73] and Retico et al. [28] implemented a histogram-based thresholding to segregate the lung region from the neighbouring structure. Thresholding, region growing, and anatomical-based knowledge was included in Antonelli et al. [74] to isolate the lung region from other structures. Gurcan et al. [22] proposed multiple thresholding for curvature indention identification. Intensity histogram analysis and morphological opening were implemented by Kubo et al. [31] to extract pulmonary nodules from the background. Morphological operators, holes filling, and reconstruction of the convex hull of the lung border were utilised by Garnavi et al. [20]. Kim et al. [21] used binarisation and labelling for lung tissue segmentation. Initial gray-level thresholding followed by morphological operations, 3D region growing, and connection cost were employed by Fetita et al. [48] to separate three classes of lung nodules. Thresholding and flood-filling were proposed by Pu et al. [25] to retain the lung region in the image slice. Adaptive border marching was applied to avoid over segmentation.

123

S. L. A. Lee et al.

Sun et al. [49] applied the mean shift analysis and region growing to extract nodule from image slices. Okada et al. [59, 75] utilised a mean shift normalised gradient as an alternative method to intensity thresholding. According to Diciotti et al. [3], segmentation algorithms should be evaluated on a large public databases with a well-defined ground truth for verification. Several of the existing studies utilised private databases. Therefore, a performance comparison between various methods is thus limited. 2.4 Nodule detection Lung nodule detection refers to the process of determining whether nodule patterns are present in the image, and identifying the location of the nodules (see Fig. 6). Some existing nodule detection systems do not include this component in their structure. They only rely on the output of the lung segmentation component. On the other hand, other systems employ both components to detect nodules more accurately. In this case, the nodule detection component is used to refine the output of the lung segmentation component. Finally, a few existing systems bypass the lung segmentation component and employ only the nodule detection component. There are a number of different approaches that can be used to realise this component. The most widely used approach is detection by classification. Other approaches including template matching and clustering are also reported in the literature. Li et al. [46] proposed an automated rule-based classifier to classify nodules and non-nodules. Also, Kostis et al. [2] characterised the malignancy status of the nodule through a rule-based approach. A linear discriminant analysis (LDA) classifier was implemented by Kawata et al. [29,53], which achieved the Az value of 0.97. Takizawa et al. [76] described a classifier based on LDA and 3D lattice system to map the relationship between nodule and image features. An improvement from the Az value of 0.918 to 0.931 was achieved. The LDA classifier was also employed by Matsumoto et al. [40] using eight features to determine the intermediate candidate nodules. If multiple true detections corresponding to

Fig. 6 Sample detected nodules in a lung image


the same nodule were detected, the overall nodule was only counted as one. A knowledge-base of anatomical structure and a rule-base of area, diameter, circularity, mean value of gray level, and smoothness of the region of interest were used by Jia et al. [50] and Fukano et al. [44] to determine nodule patterns. Armato et al. [77] employed the LDA classifier to classify the nodule candidate produces by multiple thresholding. Aoyama et al. [52] utilised the LDA classifier based on three gray-level features, two edge-based features, a morphological feature, and two clinical features from inside and outside of the nodule region to discriminate between malignant and benign nodules. Kim et al. [21] classified the ground glass opacity nodule using LDA based on the Mahalanobis distance distribution. A texture-based classifier was developed by Kim et al. [54] which classified the true nodules from the candidate nodules based on features, such as size, shape, average, standard deviation, and correlation coefficient. Artificial neural networks (ANN) were employed by Awai et al. [43] and Arimura et al. [30] for detection. A cellular neural network was presented by Zhang et al. [78] and a voxel-based neural approach was utilised by Gori et al. [27] and Retico et al. [28] to detect nodules. Linear and non-linear k-NN regressions, and support vector machine (SVM) regression were performed by Ginneken [79] to predict the unseen nodule feature vector. Zhao et al. [38] applied boosting of the k-NN classifier to estimate the probability density function of the intensity value of the trained ground glass opacity nodules. A multiple classifier was included in the work of Klik et al. [70] to classify the benign sub-pleural nodules. The classifiers were k-NN, Parzen, LDA, and quadratic discriminant. LDA and the quadratic discriminant classifiers performed slightly better than their counterparts. LDA, k-NN, quadratic discriminant, and SVM classifiers were employed by Sluimer et al. [32] to conduct an extensive evaluation. All classifiers performed well and the k-NN was chosen due to its slightly better results than the other counterparts. Shah and McNitt-Gray [66,67] employed the logistic regression, quadratic, and LDA to discriminate between malignant and benign patterns. It was reported that the logistic regression classifier recorded the highest classification performance compared to its counterparts. On the other hand, LDA recorded the lowest classification performance. The SVM classifier was used by Dehmeshki et al. [80], Zhao et al. [81], and Korfiatis et al. [19] to optimise the identification of nodule and non-nodule patterns. A massive training ANN was developed by Suzuki et al. [82–84] to detect nodule patterns, such as benign and malignant. Fuzzy rules were developed by Brown et al. [85] and Dehmeshki et al. [69] to identify the nodule patterns from within the image slices. Mean-shift clustering was utilised by Sun et al. [49] to find the nodule regions in feature space.

155

This non-parametric clustering technique is an unsupervised method which does not require previous knowledge of the cluster labels. A two-level convolution neural network was proposed in Lin et al. [86]. Lin and Yan [23] and Lin et al. [24] combined fuzzy logic and neural networks for lung nodule detection and reported that the combination was superior to rule-base, convolution neural network, and genetic algorithm template matching approaches. Also, Antonelli et al. [74] applied feed forward four-layer fuzzy neural networks which employed fuzzy rules in which the structure of the rules was learnt during the learning phase of the classifier. A nearest cluster method was used by Ezoe et al. [41] and Tanino et al. [42] to classify the previously segmented nodules candidate. An expectation maximization (EM) classifier was trained by Xu et al. [51] and effectively highlighted nodule areas. Pereira et al. [87] introduced a multi layer Perceptron classifier for nodule classification. A SVM classifier was trained by Boroczky et al. [88] on the subset of optimal features to classify the candidate nodules as nodule or non-nodule. A Bayesian classifier was trained by McCulloch et al. [89] to classify the nodule and non-nodule patterns. Template matching was developed by Wang et al. [73] and El-Baz et al. [71] to estimate the similarity between the regions adjacent to the voxels and the nodule template. Combination of genetic algorithm and template matching was initiated by Lee et al. [64,65,90] and Dehmeshki et al. [91] based on the shape features calculated at voxels and merged with global nodule intensity distribution. Originally, Ozekes et al. [16] applied a two-step classification based on location change measurement and template matching. Further improvement on the detection rate through template matching with fuzzy rule thresholding was achieved by Ozekes et al. [18] on candidate nodules. A 3D prismatic nodule template was constructed by Osman et al. [17] based on nodule features. 2D and 3D deformable template matching were developed by Farag et al. [36,37] to detect the geometry and gray level distribution within the nodules. Kawata et al. [53] proposed a linear discriminant classification boosted by k-means clustering using malignant and benign pulmonary nodule datasets based on topological histogram features. Gurcan et al. [22] used curvature analysis and candidate detection scheme using k-means clustering, rule-based followed by LDA. Ochs et al. [15] employed AdaBoost which was trained on a number of weak classifiers to simultaneously select and merge features from the feature set throughout the training of each independent classifier. Kouzani et al. [92] developed an ensemble-based random forests learner that grew many classification trees for nodule detection. Each tree produced a classification decision, and an integrated output was calculated. Lee et al. [93] developed an ensemble classification aided by clustering approach to improve the lung nodule classification performance.

123

156

S. L. A. Lee et al.

Fig. 7 False positives reduction: (left) input candidate nodules, (right) output nodules, top, and false positives, bottom, images

2.5 False positives reduction False positives reduction refers to the process of further eliminating the false positives from the output of nodule detection or lung segmentation components (see Fig. 7). It aims to achieve maximum sensitivity or true positive rate. Not all existing works incorporate a false positives reduction component in their structure. A LDA classifier was employed by Gurcan et al. [22] and Armato et al. [94] to reduce the false positives produced by a rule-based classifier. A new feature with 3D gradient field was added to the LDA classifier by Ge et al. [95] to improve the false positives of Gurcan et al. [22]. The Az value increased from 0.91 to 0.93 after applying the new feature. New features relative to the position of the anatomical structure of the lung were utilised by Saita et al. [96] to reduce the false positive rate in Oda et al.’s [35] detection algorithm. A 100% detection rate with 2.6 FPs/scan was achieved in Saita et al. [96] compared to 59% detection rate with 19.2 FPs/scan of Oda et al. [35]. Ridge detection and model matching of vessels and nodules were proposed by Fukano et al. [44] as important criteria to reduce false positives. A principal component-based clustering was employed by Tanino et al. [42] utilising the Mahalanobis distance discriminate function to distinguish the suspicious regions as normal or abnormal. Comparing with Ezoe et al.’s [41] original system, they achieved 100% detection with only 39 FPs/scan. Arimura et al. [30] developed a false positives reduction method using a rule-based classifier that localises image features followed by a multiple massive training ANN for various types of false positives. A convolution neural network was implemented by Lin et al. [86] to reduce false positives. Li et al. [46] employed a rule-based classifier with composite features to eliminate false positives. Position-dependent threshold on size was considered by Awai et al. [43] to determine the likeliness of false nodules among nodule candidates. Sphericity test was implemented by Chang et al. [39] and a multiple surface attachment filter was utilised by Enquobahrie et al. [97] to reduce false positives. Jia et al.

123

[50] and Zhao et al. [98,99] applied a knowledge-base and a rule-base to eliminate false nodule candidates. A prediction system based on the confidence level was devised by Takizawa et al. [76] to predict pulmonary nodules and false positives in lung image slices. Simple rule-based filtering was used to eliminate the false positives by Dehmeshki et al. [91]. Elimination of false positives was carried out by Osman et al. [17] on higher density values than the neighbouring nodules. A Bayesian classifier with two gray levels and a shape feature was included in the work of Farag et al. [36,37] to reduce false positives. The genetic algorithm was employed by Boroczky et al. [88] and Zhao et al. [81] to identify a subset of features to be included in the training phase of a SVM classifier that reduced false positives and retained all true positives. Incorporating the genetic algorithm and the template matching by Lee et al. [64] reduced false positives from a previous study in ref. [90]. Further improvement was recorded when Lee et al. [65] added five features and threshold values to further eradicate the false positives. SVM-based morphological features were used to reduce false positives in Korfiatis et al. [19]. A voxel-based neural approach analysed the 3D neighbourhood of the region of interest voxels to determine the likelihood of false positives in Gori et al. [27] and Retico et al. [28]. Clustering non-nodule patterns using k-mean were implemented by Dehmeshki et al. [80] through the Gaussian mixture model which reduced false positives. Lin et al. [24] eliminated false positives by determining the relationship between the location and the features of the consecutive slices. Suzuki et al. [83] carried out false positives reduction based on voxels. This was applied to false positives that could not be removed by a first false positives reduction method based on features. 3 Performance comparison of existing nodule detection methods A review of the existing nodule detection methods reveals that the main bottleneck in comparing the results of the


published works is the difference in the parameters considered in forming the methods including the properties of the training and test datasets, performance evaluation methods, and characteristics of the targeted nodule groups. We have constructed a table (see Table 1) to facilitate a comparison of the performance of several popular existing nodule detection systems. For each reported system, the table gives such information as authors, publication year, data used, and performance results. For the data used, we have included the number of scans, slices, and nodules as well as the type of nodules from the published work wherever given. For the performance results, we have included the sensitivity, specificity, false positives, false negatives, and Az from the published work wherever reported. The existing methods have been found to have various structures. Each structure involves a number of algorithmic components as well as their specific inter-relationships. The paper presented a generic structure to represent the existing methods. Considering the generic structure that highlighted the internal components of the existing systems, and the review of the algorithmic realisation of the components, the existing systems can be grouped based on their detection principle: segmentation detection, classification detection, segmentation-template detection, and segmentation–classification detection. This grouping basically represents the method by which the nodule patterns are detected in the preprocessed image. For example, in the segmentation detection, only a segmentation algorithm is used to extract nodule patterns out of the pre-processed image. On the other hand, in the hybrid segmentation–classification detection, extraction of nodule patterns is carried out first through a segmentation algorithm and the results are then refined by a classification approach. Segmentation detection methods appear in [3,12,20,24, 48,59,62,75,101]. Fetita et al. [48] reported 90%-98% sensitivity and 85–97% specificity on 10 cases consisting of 300 nodules. They described that their system was superior to conventional rule/knowledge-based, fuzzy clustering, k-means, dynamic programming, neural networks, template matching, and mathematical morphology algorithms. On the other hand, the work by Goo et al. [101] produced 65% sensitivity and 422 false positives when evaluated on 50 scans with 52 nodules. Classification detection methods appear in [15,32,39,41, 42,45,76,82,83,87]. According to [82,83], segmentation methods do not perform well because they produce incorrect segmentation for complex patterns, such as juxtavascular and subtle ground glass opacity. Whilst a sensitivity of 100% and 0.88 FPs/scan were reported by Chang et al. [39], a small number of scans and nodules used in their study. AdaBoost by Ochs et al. [15] who reported Az of 0.945 for 29 scans, and multiple ANN by Suzuki et al. [83] who achieved 94% (58/62 nodules) sensitivity with a false positive rate of 5

157

(161/32) per patient for a database containing 62 nodules in 32 patients, demonstrated good performances. On the other hand, the work by Pereira et al. [87] produced 77.71% sensitivity and 87.18% specificity for 154 nodules. The segmentation-template detection methods appear in [16–18,64,65,71,73,90,91]. Osman et al. [17], Wang et al. [73], and Ozekes et al. [18] achieved 100% sensitivity, for 6 to 47 nodule samples that were included in their studies. The work by Ozekes et al. [16] used 276 slices containing 153 nodules and produced 95% sensitivity and 1.17 FPs/slice for 20 pixels template. On the other hand, Lee et al. [90] achieved 67% sensitivity and 10 FPs/slice for 11 cases consisting of 67 nodules. The segmentation–classification detection methods appear in [21,23,24,28,30,31,34,35,43,44,46,51–54,61,63,67, 69,74,78,80,81,85,89,95,96,98–100,102,106,107]. The sensitivity reported varies from 59% by [35] to 100% by [38,42,49,74,80,88,96]. The classification aided by clustering approach by Kawata et al. [53] used a large dataset with 210 malignant and benign nodules and obtained Az of 0.97. On the other hand, the work of Oda et al. [35] produced 59% sensitivity and 19.2 FPs/scan for 33 cases with 57 nodules. Whilst some methods have demonstrated 100% sensitivity, they may not necessarily be the best methods among the available methods. This is due to the fact that sometimes only a small dataset of nodule patterns were used in the evaluation of their systems. Overall, a number of methods can be regarded as good performers. Kawata et al. [53] used a data of 210 cases and achieved Az of 0.97. Ensemble-based classification by Suzuki et al. [83] and Ochs et al. [15] produced good results. A massive training was involved in Suzuki et al. [83] multiple ANN. Ochs et al.’s [15] study does not involve automatic detection, but only classification of regions highlighted by multiple radiologists. As can be seen from Table 1, not every published work specifically reports the type of nodules used in the work. However, a number of publications state the nodule type used as well as the associated detection performance. Considering the nodule patterns associated with different nodule categories, some types of nodules, e.g. ground glass opacity, are more difficult to detect than others. Therefore, a direct comparison of the detection performances without taking into account the type of nodule used may not be very accurate. Considering the results shown in Table 1, the works by Li et al. [46] who tackled the ground glass opacity and Suzuki [84] who dealt with both the juxta-pleural and the ground glass opacity nodules achieved good sensitivities of 86 and 84%, with 6.6 FPs/scan and 0.50 FPs/section, respectively. It can be observed from the performance results reported in the literature [4–8] that lung nodule detection systems employing multi-stage algorithms for extraction of nodule patterns have performed well. For example, hybrid segmentation-classification detection methods have demonstrated

123

158

S. L. A. Lee et al.

Table 1 Performance comparison of lung nodule detection methods Author

Data

Performance

Author

Lee et al. [90]

Scans: 11

Sensitivity: 67.0%

Korfiatis et al. [19]

Nodules: 67

FP: 10.00 FPs/slice

Data

Performance

Scans: 21

Sensitivity: 81.0%

Nodules: 279

FP: 5.00 FPs/slice

Types: Juxta-pleural Kanawaza et al. [63]

Scans:450

Sensitivity: 95.0%

Gori et al. [27]

Nodules: 120 Kawata et al. [53]

Scans: 210

Az :0.97

Matsumoto et al. [40]

Armato et al. [94]

Scans: 38

Sensitivity: 80.0%

Zhao et al. [38]

Nodules: 50

Gurcan et al. [22] Kubo et al. [31]

Oda et al. [35]

Wiemker et al. [100]

Goo et al. [101] Fetita et al. [48]

Sensitivity: 85.2%

Nodules: 102

Specificity: FP: 13.50 FPs/scan Sensitivity: 90.0%

Scans:5 Nodules: 50

FP: 1.67 FPs/slice

Scans: 10

Sensitivity: 100%

FP: 1.00 FPs/slice

Nodules: 10

Specificity: FP:1.00 FP/scan

Scans: 34

Sensitivity: 84.0%

Types: Solid, ground glass opacity Scans: 38

Sensitivity: 100%

Nodules: 63

FP: 1.74 FP/slice

Nodules: 52

Specificity: 56.40%

Nodules: 155

Sensitivity: 91.0%

Boroczky et al. [88]

Scans: 70

Sensitivity: 85.0%

FP: 0.53 FPs/slice

Dehmeshki et al. [91]

Nodules: 178

FP: 14.00 FPs/scan

Types: Pleural tail, juxta-vascular, ground glass opacity Nodules: 154

Sensitivity: 77.71%

Scans: 33

Sensitivity: 59.0%

Nodules: 57

FP: 19.2FPs/scan

Pereira et al. [87]

Types: Ground glass opacity Scans: 50

Sensitivity: 86.0%

Nodules: 271

FP: 4.90FPs/scan

Nodules: 105

Sensitivity: 89.3%

Types: Isolated,juxtavascular, juxta-pleural Scans: 221

Specificity: 87.18%

Kuhnigk et al. [62]

Types: Juxta-pleural

Lin and Yan [23]

Scans: 39

Slices: 583 Nodules: 393

FP: 0.30 FPs/slice

Scans: 50

Sensitivity: 65.0%

Nodules: 52

FP: 422 FPs in total

Scans: 10

Sensitivity: 98.0%

Nodules: 300

Specificity: 97%

Tanino et al. [42]

Types: Isolated, juxta-vascular and peripheral Scans: 39

Kim et al. [54]

Types: Ground glass opacities Slices: 827

Klik et al. [70]

Scans: 16

Sensitivity: 91.4%

Az = 0.81

Nodules: 284 Ochs et al. [15]

Scans: 29

Az :0.945

Kim et al. [21]

Scans: 32

Sensitivity: 71.7%

Types: Ground glass opacity

Specificity: FP: 51.3 FPs in total

Sensitivity: 100%

Itai et al. [102]

Scans: 9

Sensitivity: 81.0% FP: 20% FN:20%

Wang et al. [73]

Types: Ground glass opacity Slices: 752

Sensitivity: 100%

Nodules: 47

FP: 21 FPs in total

Slices: 160

Sensitivity: 100%

FP: 39.00 FPs/scan Sensitivity: 96.0%

Nodules: 42 Types: Juxta-pleural Aoyama et al. [52] Awai et al. [43] Dehmeshki et al. [80]

Nodules: 489 Scans: 82

Sensitivity: 80.0%

Nodules: 78

FP: 38.7 FPs/scan

Scans: 47

Sensitivity: 100%

Nodules: 85

FP: 0.27 FPs/slice

Types: Solid

123

Az : 0.85

Osman et al. [17]

Nodules: 6

FP: 0.46 FPs/slice

Jia et al. [50]

Scans: 8

Sensitivity: 95.0%

Slices: 943

FP: 0.91 FPs/slice

Nie et al. [103]

Nodules: 39

Sensitivity: 89.0%


159

Table 1 continued Author Arimura et al. [30] Chang et al. [39]

Lee et al. [65]

Data

Performance

Author Ozekes et al. [18]

Scans: 106

Sensitivity: 83.0%

Nodules: 109

FP: 5.80 FPs/scan

Scans: 8

Sensitivity: 100%

Nodules: 62

Data

Performance

Scans: 16

Sensitivity: 100%

Nodules: 16

FP: 13.375 FPs/scan

Li et al. [46]

Scans: 117

Sensitivity: 86%

FP: 0.88 FPs/scan

Nodules: 153

FP: 6.6 FPs/scan

Scans: 20

Sensitivity: 72.4%

Retico et al. [28]

Types: Solid, ground glass opacity Scans: 39

Sensitivity: 85.0%

Nodules: 98

FP: 5.50 FPs/scan

Nodules: 102

FP: 13.00 FPs/scan

Types: Juxta-pleural Saita et al. [96]

Scans: 12

Sensitivity: 100%

Diciotti et al. [3]

Nodules: 145

Lee et al. [93]

Types: Well-circumscribed, juxta-vascular, juxta-pleural, pleural tail Scans: 32

FP: 2.60FPs/scan

Farag et al. [36,37] Suzuki et al. [83]

Shah et al. [67]

Scans: 200

Sensitivity: 82.3%

Nodules: 130

FP: 12.00 FPs in total

Scans: 32

Sensitivity: 94%

Nodules: 62

FP: 5 per scan

Scans: 35

Az : 0.92

Sensitivity: 86.32% for Italung-CT and 83.33% for the LIDC

Sensitivity: 96.67%

Nodules: 1203 Suzuki [84]

Nodules: 69

Sensitivity: 84.0% FP: 0.50 FPs/section

Murphy et al. [104]

Types: Juxtapleural and groundglass opacity Scans: 813(A)

Nodules: 35

541 (B)

72.4%(B)

Types: Solid and ground glass opacity

541 (C)

77.7%(C) FP:4.20 FPs/scans (A) 4.00 FPs/scans (B)

Scans: 59

Az : 0.83 +/- 0.04

Sensitivity: 80.0%(A)

4.20 FPs/ scans (C) Lin et al. [24]

Scans: 29

Sensitivity: 89.3%

Nodules: 393

FP: 0.21 FPs/slice

Way et al. [105]

Nodules: 96

Types: Juxta-pleural, pleural tail, juxta-vascular

better performances than those of segmentation only detection methods. It can be stated that detection of nodule patterns within lung images is a challenging task. Tackling this task by using a single algorithm, such as a standard segmentation approach cannot deliver maximum true positives and minimum false positives at the same time. However, a multistage approach employs different algorithms each tackling one aspect of the detection task, and together creating a platform for delivering the expected performance targets. The ability to choose the right set of features for the false positives reduction is still being investigated. Choosing too many features will cause high computational time and over-fitting while selecting too few features will make the classifier inefficient. It can be concluded that the performance of several existing systems can be considered supe-

rior [4]. The exhibited performances can be further improved by employing the hybrid approach and also developing and utilising proper false positives reduction components. A well performing automated lung nodule detection system can help expert radiologist in detecting lung cancer, and therefore saving lives.

4 Conclusions This paper presented a study of the existing lung nodule detection methods. It gave a generic structure for lung nodule detection methods. The algorithms used to realise various components of the existing systems were described. The review of the existing methods demonstrates that the

123

160

bottleneck in comparing the results of the published work is the difference in the parameters considered in forming the methods including the properties of training and test datasets and performance evaluation methods. The paper formulated a comparison of the performance of the existing reported approaches. Overall, systems that employ multi-stage detection algorithms have been more popular and have demonstrated better performances. It can be concluded that the performance of several existing systems have been superior. The exhibited performances can be improved by employing proper false positives reduction components. References 1. Austin, J.H., Mueller, N.L., Friedman, P.J.: Glossary of terms for CT of the lungs: recommendations of the nomenclature. Comm. Fleischner Soc. Radiol. 200, 327–331 (1996) 2. Kostis, W.J., Reeves, A.P., Yankelevitz, D.F., et al. : Threedimensional segmentation and growth-rate estimation of small pulmonary nodules in helical CT images. IEEE Trans. Medical Imaging 22, 1259–1274 (2003) 3. Diciotti, S., Picozzi, G., Falchini, M., et al. : 3D segmentation algorithm of small lung nodules in spiral CT images. IEEE Trans. Inf. Technol. Biomedical 12, 7–19 (2008) 4. Li, Q.: Recent progress in computer-aided diagnosis of lung nodules on thin-section CT. Comput. Med. Imaging Graph. 31, 248– 257 (2007) 5. Sluimer, I.C., Schilham, A., Prokop, M., et al.: Computer analysis of computed tomography scans of the lung: a survey. IEEE Trans. Medical Imaging 25, 385–405 (2006) 6. Jeong, Y.J., Yi, C.A., Lee, K.S.: Solitary pulmonary nodules: detection, characterization, and guidance for further diagnostic workup and treatment. AJR 188 (2007) 7. Parrish, F.J.: Volume CT: state-of-the-art reporting. AJR 189, 528– 534 (2007) 8. Doi, K.: Computer-aided diagnosis in medical imaging: historical review, current status and future potential. Comput. Med. Imaging Graph. 31, 198–211 (2007) 9. Marten, K., Engelke, C.: Computer-aided detection and automated CT volumetry of pulmonary nodules. Eur Radiol 17, 888– 901 (2007) 10. ELCAP public lung image database. Vision and Image Analysis Group (VIA) and International Early Lung Cancer Action Program (I-ELCAP) Labs, Cornell University. http://www.via. cornell.edu/lungdb.html (2007) 11. Public lung database to address drug response. Vision and Image Analysis Group (VIA) and International Early Lung Cancer Action Program (I-ELCAP) Labs, Cornell University. http:// www.via.cornell.edu/crpf.html (2007) 12. Lung Imaging Database Consortium (LIDC). http://imaging. cancer.gov/programsandresources/InformationSystems/LIDC. 13. Medical image database. MedPix. http://rad.usuhs.edu/medpix/ index.html (2007) 14. Wang, J., Engelmann, R., Li, Q.: Segmentation of pulmonary nodules in three-dimensional CT images by use of a spiral-scanning technique. Med. Phys. 34, 4678–4689 (2007) 15. Ochs, R.A., Goldin, J.G., Fereidoun, A., et al.: Automated classification of lung bronchovascular anatomy in CT using Adaboost. Med. Image Anal. 11, 315–324 (2007) 16. Ozekes S., Camurcu A.Y.: Automatic lung nodule detection using template matching. In: Yakhno T. E. N. (ed.) Lecture Notes in Computer Science, vol. 4243. pp. 247–253 (2006)

123

S. L. A. Lee et al. 17. Osman, O., Ozekes, S., Ucan, O.N.: Lung nodule diagnosis using 3D template matching. Comput. Biol. Med. 37, 1167–1172 (2007) 18. Ozekes, S., Osman, O., Ucan, O.N.: Nodule detection in a lung region that’s segmented with using genetic cellular neural networks and 3D template matching with fuzzy rule based thresholding. Korean J. Radiol. 9, 1–9 (2008) 19. Korfiatis, P., Kalogeropoulou, C., Costaridou, L.: Computer aided detection of lung nodules in multislice computed tomography. In: Proceedings of the International Special Topic Conference on Information Technology in Biomedicine (IEEE-ITAB 2006), p. 4. Ioannina, Epirus, Greece (2006) 20. Garnavi, R., Baraani-Dastjerdi, A., Abrishami Moghaddam, H., et al.: A new segmentation method for lung HRCT images. In: Lovell, B.C., Maeder, A.J., Caelli, T., Ourselin, S. (eds.): Proceedings of the Digital Imaging Computing: Techniques and Applications, p. 8. IEEE CS Press, Cairns Convention Centre, Brisbane, Australia (2005) 21. Kim, H., Nakashima, T., Itai, Y., et al.: Automatic detection of ground glass opacity from the thoracic MDCT images by using density features. In: International Conference on Control, Automation and Systems, pp. 1274–1277. IEEE Xplore, COEX, Seoul, Korea (2007) 22. Gurcan, M.N., Sahiner, B., Petrick, N., et al.: Lung nodule detection on thoracic computed tomography images: preliminary evaluation of a computer-aided diagnosis system. Med. Phys. 29, 2552–2558 (2002) 23. Lin, D.T., Yan, C.R.: Lung nodules identification rules extraction with neural fuzzy network. In: Wang, L., Rajapakse, J.C., Fukushima, K., Lee, S.Y., Yao, X. (eds.): 9th International Conference of Information Processing (ICONIP), IEEE, Singapore (2002) 24. Lin, D.T., Yan, C.R., Chen, W.T.: Autonomous detection of pulmonary nodules on CT images with a neural network-based fuzzy system. Comput. Med. Imaging Graph. 29, 447–458 (2005) 25. Pu, J., Roos, J., Yi, C.A., et al.: Adaptive border marching algorithm: automatic lung segmentation on chest CT images. Comput. Med. Imaging Graph. 32, 452–462 (2008) 26. Wei, G.Q., Fan, L., Qian, J.: Automatic detection of nodules attached to vessels in lung CT by volume projection analysis. Medical Image Computing and Computer-assisted Intervention— MICCAI, vol. 2488, pp. 746–752. Springer, Berlin (2002) 27. Gori, I., Bellotti, R., Cerello, P., et al.: Lung nodule detection in screening computed tomography. IEEE Nuclear Science Symposium Conference Record. IEEE (2006) 28. Retico, A., Delogu, P., Fantacci, M.E., et al.: Lung nodule detection in low-dose and thin-slice computed tomography. Comput. Biol. Med. 38, 525–534 (2008) 29. Kawata, Y., Niki, N., Ohmatsu, H., et al. : Quantitative surface characterization of pulmonary nodules based on thin-section CT images. IEEE Trans. Nuclear Sci. 45, 2132–2138 (1998) 30. Arimura, H., Katsuragawa, S., Suzuki, K., et al. : Computerized scheme for automated detection of lung nodules in low-dose computed tomography images for lung cancer screening. Acad. Radiol. 11, 617–629 (2004) 31. Kubo, M., Kubota, K., Yamada, N., et al.: A CAD system for lung cancer based on low dose single-slice CT image. In: Proc. SPIE, vol. 4684, pp. 1672–1680 (2002) 32. Sluimer, I.C., van Waes, P.F., Vierever, M.A., et al. : Computer-aided diagnosis in high resolution CT of the lungs. Med. Phys. 30, 3081–3090 (2003) 33. Bae, K.T., Kim, J.S., Na, Y.H., et al.: Pulmonary nodules: automated detection on CT images with morphologic matching algorithm—preliminary results. Radiology 236, 286–294 (2005) 34. Paik, D.S., Beaulieu, C.F., Rubin, G.D., et al. : Surface normal overlap: a computer-aided detection algorithm with application


35.

36.

37.

38.

39.

40.

41.

42.

43.

44.

45.

46.

47.

48.

49.

50.

51.

to colonic polyps and lung nodules in helical CT. IEEE Trans. Med. Imaging 23, 661–675 (2004) Oda, T., Kubo, M., Kawata, Y., et al.: A detection algorithm of lung cancer candidate nodules on multi-slice CT images. In: Proc. of SPIE, vol. 4684. pp. 1354–1361 (2002) Farag, A., El-Baz, A., Gimel’farb, G.G., et al.: Automatic detection and recognition of lung abnormalities in helical CT images using deformable templates. In: Barillot, C., Haynor, D.R., Hellier, P. (eds.): Medical image Computing and Computerassisted Intervention—MICCAI, vol. 3217, pp. 856–864. Springer, Berlin (2004) Farag, A.A., El-Baz, A., Gimel’farb, G., et al.: Detection and recognition of lung abnormalities using deformable templates. In: Proceeding of 17th International Conference on Pattern Recognition (CVIP), pp. 23–26, August 2004. IEEE (2004) Zhao, J., Chang, S., Metaxas, D.N., et al.: An automatic method for ground glass opacity nodule detection and segmentation from CT studies. In: Proceedings of 28th IEEE EMBS Conference, pp. 3062–3065. IEEE Xplore, New York, USA (2006) Chang, S., Emoto, H., Metaxas, D.N., et al.: Pulmonary micronodule detection from 3-d chest CT. Medical image computing and computer-assisted intervention, vol. 3217, pp. 821–828. Springer, Berlin (2004) Matsumoto, S., Kundel, H.L., Gee, J.C., et al.: Pulmonary nodule detection in CT images with quantized convergence index filter. Medi. Image Anal. 10, 343–352 (2006) Ezoe, T., Takizawa, H., Yamamoto, S., et al.: An automatic detection method of lung cancers including ground glass opacities from chest X-ray CT images. In: Proc. of SPIE, vol. 4684, pp. 1672– 1680 (2002) Tanino, M., Takizawa, H., Yamamoto, S., et al.: A detection method of ground glass opacities in chest X-ray CT images using automatic clustering techniques. In: Proc. of SPIE, vol. 5032, pp. 1728–1737 (2003) Awai, K., Murao, K., Ozawa, A., et al.: Pulmonary nodules at chest CT: effect of computer-aided diagnosis on radiologists’ detection performance. Radiology 230, 347–352 (2004) Fukano, G., Takizawa, H., Shigemoto, K., et al.: Recognition method of lung nodules using blood vessel extraction techniques and 3-D object models. In: Proc. of SPIE, vol. 5032, pp. 190–196 (2003) Li, Q., Doi, K.: New selective enhancement filter and its application for significant improvement of nodule detection on computed tomography. In: Proc. of SPIE, vol. 5370, pp. 1–9 (2004) Li, Q., Li, F., Doi, K.: Computerized detection of lung nodules in thin-section CT images by use of selective enhancement filters and an automated rule-based classifier. Acad. Radiol. 15, 165– 175 (2008) Yu, Y., Zhao, H.: Enhancement filter for computer-aided detection of pulmonary nodules on thoracic CT images. In: Proceeding of the Sixth International Conference on Intelligent Systems Design and Applications. IEEE Xplore (2006) Fetita, C.I., Prêteux, F.J., Beigelman-Aubry, C., et al.: 3-D automated lung nodule segmentation in HRCT. Lecture Notes in Computer Science, vol. 2878/2003, pp. 626–634. Springer, Berlin (2003) Sun, S.-S., Li, H., Hou, X.-R., et al.: Automatic segmentation of pulmonary nodules in CT images. In: 1st International Conference on Bioinformatics and Biomedical Engineering (ICBBE), pp. 790–793. IEEE (2007) Jia, T., Zhao, D.-Z., Yang, J.-Z., et al.: Automated detection of pulmonary nodules in HRCT images. In: 1st International Conference on Bioinformatics and Biomedical Engineering (ICBBE), pp. 833–836. IEEE (2007) Xu, N., Ahuja, N., Bansal, R.: Automated lung nodule segmentation using dynamic programming and em-based classification.

161

52.

53.

54. 55.

56.

57.

58.

59.

60.

61.

62.

63.

64.

65.

66.

67.

68.

In: Sonka, M., Fitzpatrick, J.M. (eds.) Proc. of SPIE, vol. 4684, pp. 666–676 (2002) Aoyama, M., Li, Q., Katsuragawa, S., et al.: Computerized scheme for determination of the likelihood measure of malignancy for pulmonary nodules on low-dose CT images. Med. Phys. 30, 387– 394 (2003) Kawata, Y., Niki, N., Ohmatsu, H., et al.: Hybrid classification approach of malignant and benign pulmonary nodules based on topological and histogram features. Medical Image Computing and Computer-assisted Intervention—MICCAI 2000, vol. 1935/2000, pp. 297–306. LNCS Springer Berlin/Heidelberg, Berlin (2000) Kim, D.Y., Kim, J.H., Noh, S.M., et al.: Pulmonary nodule detection using chest CT images. Acta. Radiol. 44, 252–257 (2003) Bellotti, R., De Carlo, F., Gargano, G., et al.: A CAD system for nodule detection in low-dose lung cts based region growing and active contour models. Med. Phys. 34, 4901–4910 (2007) Itai, Y., Hyoungseop, K., Ishida, T., et al.: A segmentation method of lung areas by using snakes and automatic detection of abnormal shadow on the areas. Int. J. Innov. Comput. Inf. Control 3, 277– 284 (2007) Zhao, B., Yankelevitz, D., Reeves, A., et al. : Twodimensional multi-criterion segmentation of pulmonary nodules on helical CT images. Med. Phys. 26, 889–895 (1999) Ko, J.P., Rusinek, H., Jacobs, E., et al.: Volume quantitation of small pulmonary nodules on low dose chest CT: a phantom study. In: Radiological Society of North America 87th Scientific Assembly and Annual Meeting, Chicago (2001) Okada, K., Comaniciu, D., Krishnan, A.: Robust anisotropic Gaussian fitting for columetric characterization of pulmonary nodules in multislice ct. IEEE Trans. Med. Imaging 24, 409– 423 (2005) Fan, L., Qian, J., Odry, B.L., et al.: Automatic segmentation of pulmonary nodules by using dynamic 3-d cross-correlation for interactive CAD systems. In: Proc. SPIE Med. Imag, vol. 4684, pp. 1362–1369 (2002) Enquobahrie, A.A., Reeves, A.P., Yankelevitz, D.F., et al.: Automated detection of pulmonary nodules from whole lung helical CT scans: performance comparison for isolated and attached nodules. In: Proc. of SPIE, vol. 5370. pp. 791–800, columetric (2004) Kuhnigk, J.-M., Dicken, V., Bornemann, L., et al.: Morphological segmentation and partial volume analysis for volumetry of solid pulmonary lesions in thoracic CT scans. IEEE Trans. Med. Imaging 25, 417–434 (2006) Kanazawa, K., Kawata, Y., Niki, Y., et al.: Computer-aided diagnosis for pulmonary nodules based on helical CT images. Comput. Med. Imaging Graph. 22, 157–167 (1998) Lee, Y., Hara, T., Fujita, H., et al.: Automated detection of pulmonary nodules in helical CT images based on an improved template-matching technique. IEEE Trans. Med. Imaging 20, 595– 604 (2001) Lee, Y., Tsai, D.Y., Hara, T., et al.: Improvement in automated detection of pulmonary nodules on helical X-ray CT images. In: Proc. of SPIE, vol. 5370, pp. 824–832. (2004) McNitt-Gary, M.F., Hart, E., Wyckoff, M.N., et al. : A pattern classification approach to characterizing solitary pulmonary nodules imaged on high resolution CT: preliminary results. Med. Phys. 26, 880–888 (1999) Shah, S.K., McNitt-Gray, M.F., Rogers, S.R., et al. : Computer aided characterization of solitary pulmonary nodules using volumetric and contrast enhancement features. Acad. Radiol. 12, 1310–1319 (2005) Udupa, J.K., Saha, P.K., Lotufo, R.A.: Relative fuzzy connectedness and object definition: theory, algorithms, and applications in image segmentation. IEEE Trans. Pattern Anal. Mach. Intell. 24, 1485–1500 (2002)

123

162 69. Dehmeshki, J., Amin, H., Valdivieso, M., et al. : Segmentation of pulmonary nodules in thoracic CT scans: a region growing approach. IEEE Trans. Medi. Imaging 27, 467–480 (2008) 70. Klik, M.A.J., Rikxoort, E.M.v., Peters, J.F., et al.: Improved classification of pulmonary nodules by automated detection of benign subpleural lymph nodes. ISBI (2006) 71. El-Baz, A., Gimel’farb, G., Falk, R., et al.: A new CAD system for early diagnosis of detected lung nodules. IEEE Int. Conf. Image Process. 2:II-461–II-464 (2007) 72. Giger, M.L., Bae, K.T., MacMahon, H.: Computerized detection of pulmonary nodules in computed tomography images. Invest. Radiol. 29, 459–465 (1994) 73. Wang, P., DeNuzio, A., Okunieff, P., et al. : Lung metastases detection in CT images using 3D template matching. Med. Phys. 34, 915–922 (2007) 74. Antonelli, M., Frosini, G., Lazzerini, B., et al.: A CAD system for lung nodule detection based on an anatomical model and a fuzzy neural network. North American Fuzzy Information Processing Society, vol. pp. 448–453. IEEE Xplore (2006) 75. Okada, K., Comaniciu, D., Krishnan, A.: Robust 3D segmentation of pulmonary nodules in multislice CT images. Lecture Notes in Computer Science, pp. 881–889. Springer, Berlin (2004) 76. Takizawa, H., Yamamoto, S., Shiina, T.: Accuracy improvement of pulmonary nodule detection based on spatial statistical analysis of thoracic CT scans. IEICE Trans. Inf. Syst. 90-D, 1168– 1174 (2007) 77. Armato S.G., Giger M.L., Moran C.J., et al. (1999) Computerized detection of pulmonary nodules on CT scans. Radiographics 19: 1303–1311 78. Zhang, X., McLennan, G., Hoffman, E.A., et al.: Computerized detection of pulmonary nodules using cellular neural networks in CT images. In: Proc. of SPIE, vol. 5370, pp. 30–41 (2004) 79. Ginneken, B.v.: Supervised probabilistic segmentation of pulmonary nodules in CT scans. In: 9th Medical Image Computing and Computer-assisted Intervention –MICCAI Conference. Springer, Berlin (2006) 80. Dehmeshki, J., Chen, J., Casique, M.V., et al.: Classification of lung data by sampling and support vector machine. 26th In: Annual International Conference of the IEEE Engineering in Medicine and Biology Society (IEMBS ‘04), vol. 2, pp. 3194– 3197 (2004) 81. Zhao, L., Boroczky, L., Lee, K.P.: False positive reduction for lung nodule CAD using support vector machines and genetic algorithms. Int. Congress Ser. 1281, 1109–1114 (2005) 82. Suzuki, K., Li, F., Sone, S., et al. : Computer-aided diagnostic scheme for distinction between benign and malignant nodules in thoracic low-dose CT by use of massive training artificial neural network. IEEE Trans. Med. Imaging 24, 1138–1150 (2005) 83. Suzuki, K., Li, F., Li, Q., et al. : Comparison between 2D and 3D massive-training ANNs (MTANNs) in CAD for lung nodule detection on MDCT. Int. J. Comput. Assist. Radiol. Surg. 1, 354– 355 (2006) 84. Suzuki, K.: A supervised ‘lesion-enhancement’ filter by use of a massive-training artificial neural network (MTANN) in computer-aided diagnosis (CAD). Phys. Med. Biol. 54, S31–S45 (2005) 85. Brown, M.S., McNitt-Gray, M.F., Mankovich, N.J., et al.: Method for segmenting chest CT image data using an anatomical model: preliminary results. IEEE Trans. Med. Imaging 16, 828– 839 (1997) 86. Lin, J.-S., Lo, S.C.B., Hasegawa, A., et al. : Reduction of false positives in lung nodule detection using a two-level neural classification. IEEE Trans. Med. Imaging 15, 206–216 (1996) 87. Pereira, C.S., Alexandre, L.A., Mendonça, A.M., et al.: A multiclassifier approach for lung nodule classification. In: Campilho, A., Kamel, M. (eds.): Image Analysis and Recognition,

123

S. L. A. Lee et al.

88.

89.

90.

91.

92.

93.

94.

95.

96.

97.

98.

99.

100.

101.

102.

103.

104.

105.

vol. 4142/2006, pp. 612-623. Springer Berlin/Heidelberg, Berlin (2006) Boroczky, L., Zhao, L., Lee, K.P.: Feature subset selection for improving the performance of false positive reduction in lung nodule CAD. IEEE Trans. Inf. Technol. Biomed. 10, 504–511 (2006) McCulloch, C.C., Kaucic, R.A., Mendonca, P.R., et al.: Modelbased detection of lung nodules in computed tomography exams. Thoracic computer-aided diagnosis. Acad. Radiol. 11, 258– 266 (2004) Lee, Y., Hara, T., Fujita, H., et al.: Nodule detection on chest helical CT scans by using a genetic algorithm. In: Proceedings of Intelligent Information Systems, IIS ‘97., pp. 67–70. Grand Bahama Island, Bahamas (1997) Dehmeshki, J., Ye, X., Casique, M.V., et al.: A hybrid approach for automated detection of lung nodules in CT images. In: 3rd IEEE International Symposium on Biomedical Imaging: Nano to Macro, pp 506–509. IEEE (2006) Kouzani, A., Lee, S., Hu, E.J.: Lung nodules detection by ensemble classification. In: Poo, A.-N. (ed.) Proceedings of 2008 IEEE International Conference on Systems, Man, and Cybernetics (SMC 2008), pp. 324-329. IEEE Xplore, Singapore (2008) Lee, S.L.A., Kouzani, A.Z., Hu, E.J.: Random forest based lung nodule classification aided by clustering. Comput. Med. Imaging Graph. (2010) (in press) Armato S.G., Li F., Giger M.L., et al. Lung cancer: performance of automated lung nodule detection applied to cancers missed in a CT screening program. Radiology 225, 685–692 (2002) Ge, Z., Sahiner, B., Chan, H.P., et al.: Computer aided detection of lung nodules: False positive reduction using a 3-d gradient field method. Proc. of SPIE, vol. 5370. pp. 1076-1082. (2004) Saita, S., Oda, T., Kubo, M., et al.: Nodule detection algorithm based on multi-slice CT images for lung cancer screening. In: Proc. of SPIE, vol. 5370, pp. 1083–1090 (2004) Enquobahrie, A.A., Reeves, A., Yankelevitz, D.F., et al.: Automated detection of small pulmonary nodules in whole lung CT scans. Acad. Radiol. 14, 579–593 (2007) Zhao, B., Gamsu, G., Ginsberg, M.S., et al.: Automatic detection of small lung nodules on CT utilizing a local density maximum algorithm. Appl. Clin. Med. Phys. 4, 248–260 (2003) Zhao, B., Ginsberg, M.S., Lefkowitz, R.A., et al.: Application of the LDM algorithm to identify small lung nodules on low-dose MSCT scans. In: Proc. of SPIE, vol. 5370, pp. 818–823 (2004) Wiemker, R., Rogalla, P., Zwartkruis, A., et al.: Computer aided lung nodule detection on high resolution CT data. In: Proc. SPIE Conf. Medical Imaging, vol. 4684, pp. 677–688. San Diego, CA (2002) Goo, J.M., Lee, J.W., Lee, H.J., et al.: Automated lung nodule detection at low-dose CT: Preliminary experience. Korean J. Radiol. 4 (2003) Itai, Y., Hyoungseop, K., Ishida, T., et al.: A segmentation method of lung areas by using snakes and automatic detection of abnormal shadow on the areas. Int. J. Innov. Comput. Inf. Control 3, 277– 284 (2007) Nie S.-D., Li, L.-H., Chen, Z.-X.: A CI feature-based pulmonary nodule segmentation using three-domain mean shift clustering. In: International Conference on Wavelet Analysis and Pattern Recognition, pp. 223–227. IEEE Xplore, Beijing, China (2007) Murphy, K., van Ginneken, B., Schilham, A.M., et al.: A largescale evaluation of automatic pulmonary nodule detection in chest CT using local image features and k-nearest-neighbour classification. Med. Image Anal. 13, 757–770 (2009) Way, T.W., Sahiner, B., Chan, H.P., et al. : Computer-aided diagnosis of pulmonary nodules on CT scans: improvement of

Automated detection of lung nodules in CT images classification performance with nodule surface features. Med. Phys. 36, 3086–3098 (2009) 106. Armato, S.G., MacMahon, H.: Automated lung segmentation and computer-aided diagnosis for thoracic CT scans. Acad. Radiol. 11, 1011–1021 (2004) 107. Brown, M.S., Goldin, J.G., Suh, R.D., et al.: Lung micronodules: automated method for detection at thin-section CT-initial experience. Radiology 226, 256–262 (2003)

Author Biographies Shu Ling Alycia Lee received her Bachelor of Engineering degree with first class honours, and then Master of Engineering degree, both in Electronics from Deakin University in 2006 and 2010, respectively. Her research interests include biomedical image analysis.

163 Eric J. Hu received his basic professional degree of Bachelor of Engineering (Mechanical) from Zhejiang University, China in 1984. He was majored in thermal power station technologies. After graduation, he proceeded to Beijing Solar Energy Research Institute for the degree of Master of Engineering (energy technology). Two and a half years later, he started work with the same institute as a research engineer for another three years. In January 1990, Eric was granted a French government scholarship to undertake the Doctor of Engineering program in Energy Technology Division at the Asian Institute of Technology in Bangkok Thailand. He obtained the doctoral degree (D.Eng. in Energy Technology) and moved to Australia at the end of 1992. Eric worked as a Lecturer and Senior Lecturer in Thermodynamics and Fluid Mechanics at Gippsland School of Engineering, Monash University until 1999 when he joined the School of Engineering at Deakin University. He was promoted to Associate Professor in 2005 at Deakin University. He joined the School of Mechanical Engineering at the University of Adelaide, as Associate Professor in Sustainable Energy Engineering, in Feb 2009.

Abbas Z. Kouzani received his M.Eng. degree in Electrical and Electronics Engineering from the University of Adelaide, Australia in 1995, and Ph.D. degree in Electrical and Electronics Engineering from Flinders University, Australia in 1999. He was a lecturer with the School of Engineering, Deakin University, and then a Senior Lecturer with the School of Electrical Engineering and Computer Science, University of Newcastle, Australia. Currently, he is an Associate Professor with the School of Engineering, Deakin University. His research interests include intelligent micro-electro-mechanical systems.

123