Lung Segmentation and Nodule Detection in Postero Anterior Chest Radiographs

Transcription

1 Lung Segmentation and Nodule Detection in Postero Anterior Chest Radiographs Paola Campadelli, Elena Casiraghi, Simon Columbano Dipartimento di Scienze dell Informazione Università degli Studi di Milano Via Comelico, 39/ Milano, Italy 1 Abstract The use of image processing techniques and Computer Aided Diagnosis (CAD) systems has proved to be effective for the improvement of radiologists diagnosis, especially in the case of lung nodules detection. In this paper we describe a method for processing Postero Anterior chest radiographs which segments the lung field and extracts a set of nodule candidate regions characterized by low cardinality and a high sensitivity ratio. 2 Introduction In the field of medical diagnosis a wide variety of imaging techniques is currently available, such as radiography, computed tomography (CT) and magnetic resonance imaging (MRI). Although the last two are more precise and more sensitive techniques, chest radiography is still by far the most common type of procedure for the initial detection and diagnosis of lung cancer, due to its noninvasivity characteristics, radiation dose and economic considerations. Studies reported in [30] and [19] explain why chest radiograph is one of the most challenging radiograph to produce technically and to interpret diagnostically. These are the main reasons why in the last two decades a lot of research has been focused on the creation of CAD systems aimed to lung nodules detection [7], [9], [8], [17], [18], [16], [31], [10], [13], [15] [20], [22], [25], [24], [2], [35], [34], [33], [30], [11], [12], [31]. All systems are based on algorithms for the extraction of a first set of candidate regions followed by methods to select, among them, the true positives. The problem common to all the referred schemes is the high number of false positives initially extracted. Attempts to reduce this high number lead to a significant loss of true positives. Here we present a lung segmentation algorithm and a multiscale approach to enhance the visibility (also called spicularity) of the nodules; we then extract a set of candidate regions with a low cardinality and a high sensitivity ratio [1]. The enhancement procedure has proven to be effective also for extremely subtle nodules.

2 3 Materials and methods The method has been developed and tested on a standard database acquired by the Japanese Society of Radiological Technology (JSRT) and described in [23]. It is a standard database containing a total of 247 radiographs: 154 containing lung nodules of different diameter (ranging from 5 to 35 mm) and subtleties (ranging from 1 to 5, from obvious to extremely subtle ), and 93 of patients with no disease. The images were digitized with a mm pixel size, a matrix size of 2048 by 2048, and 4096 grey levels. Before processing, we have down-sampled the images to a dimension of 256 by 256 pixels: the size has been experimentally chosen in order to reduce the computational time of the algorithm without losing any details that could influence the performances of the algorithm. 4 Segmentation of the lung field The first step of an automatic system for lung nodule detection, and in general for any further analysis of chest radiographs, is the segmentation of the lung field so that the algorithms for the identification of lung nodules will be applied just to the lung area. The segmentation algorithms proposed in the literature can be grouped into: rule based systems ([3], [32], [6], [5], [4]), pixel classification methods including Neural Networks ([21], [26]) and Markov random fields ([29]), active shape models and their extensions ([27]). One of the limitations of these methods is that none of them includes in the area of interest the bottom of the chest and the region behind the heart, where lung nodules may be present. Moreover they are often based on several assumptions about the position and orientation of the thorax in the image. Our segmentation method includes in the lung area also the parts usually excluded by the methods presented in the literature and avoids all kind of assumptions such as those previously mentioned. We work with images where the chest is not always located in the central part of the image, it can be tilted and it can have structural abnormalities. To detect the vertical axis and the lung borders two different edge detection techniques are employed. The first one is based on the application of first derivatives of gaussian filters taken at 4 different orientations. The second one is based on the application of the Laplacian of Gaussian (LoG) operator at 3 different scales. The results obtained with both techniques are used by a complex edge tracking algorithm which creates a very good contour describing the lateral border of each lung. Since the edge pixels at the base of both lungs and along their central borders are not always clearly detectable we prefer to design a rough outline containing the region of interest than to lose part of the lung field. For this reason we delimit the dorsal column by simply creating two lines parallel to the vertical axis and passing by the point which is the nearest to the axis itself. As bottom boundery we use the segment connecting the two bottommost points detected by the edge tracking algorithm in the two lungs (see [Fig.1])

3 Fig. 1. Some of the results obtained We tested the segmentation algorithm both on the images of the JRST database and on 162 radiographs obtained from Niguarda Hospital. We detected 10 and 5 small errors respectively, where we consider as error a part of the lung that has not been included in the contour. On the JRST database, the performance of our method is better than that of the best method tested on the JRST database and described in [28]. 5 Enhancing the spicularity of the nodules To enhance the spicularity of nodules of different sizes and brightness we use a multiscale approach, so that also subtle and little nodules appear as visible as the obvious ones. Specifically, we produce several smoothed version of the image by convolving it with gaussian filters whose standard deviation s takes values in the range 2 12, according to the minimum and maximum possible pixel size of

4 the nodule radius. For each scale s we then subtract from the original image its smoothed version, so that we get a resulting difference image were the details visible at the scale s are enhanced. Since the distribution of gray levels in a nodule sub-image can be approximated by a gaussian, the result of subtracting to a nodule sub-image its smoothed version is usually an image with a positive peak in the central part of the nodule, and negative values in the neighborhood. Moreover the histogram of the difference image shows that most of the pixels take negative values, while on the set of positive values a peak can always be identified. We create a binary image by selecting all the pixels with a value bigger than the one corresponding to the peak. These pixels are the ones corresponding to the highest frequencies, i.e. the details, that can be identified at the scale s. Summing up all the binary images obtained at different scales s we get a final sum image (see Fig.2) whose processing is described in the next paragraph. Fig. 2. Original image and sum image - subtle nodule(top row) and extremely subtle nodule (bottom row) 6 Extracting the nodule candidates In the sum image the nodules often appear as regions with circular shape of different sizes, characterized by the highest values and surrounded by a much darker ring. Based on this observation we process this image looking for a measure which helps in selecting the pixels corresponding to the centers of nodules. To handle all the possible sizes of the nodules the procedure described below is

5 repeated for each possible radius value and all the results are then combined. Having fixed the radius R, we calculate for each pixel P = P (x, y) a coefficient P R defined as: P R =MEAN(Circle R (P )) MEAN(Ring R (P )) where Circle R (P ) is the region composed by the pixels contained in the circle of radius R and centered in P, Ring R (P ) is the region composed by the pixels in the 2-pixel-thick ring around the circle Circle R (P ), MEAN(X) is equal to the mean of the gray values of the pixels inside a generic region X. Note that the thickness of the ring is fixed to 2 for every radius. This is because what allows to identify a circular region is a darker ring surrounding it, no matter which is the thickness of the ring itself. To select the pixels which are potential nodule centers, we automatically define a threshold on the set of the coefficients P R by means of the algorithm described in [14]. For each connected region in the obtained binary image, we calculate its circularity as defined in [7], the biggest diagonal D of the minimum ellipse containing the region itself, and discard it either if the circularity is lower than 0.5 or D is bigger than 2R. The left regions correspond to the candidate nodules with radius R. Repeating the procedure for each possible radius value we obtain a set of 11 binary images, each containing a set of candidate nodules. All these images must be combined to extract a final set of candidates. Overlapping regions in two different binary images are considered to be the same one; between them we choose as representative the one with the most circular shape and discard the others. All the regions appearing in only one of the binary images are taken as candidates. With this extraction scheme we get a set of about regions on all the 247 images of the database, with an average of about 130 regions per image and only 7 true positives lost out of 153. These results have been compared with those of the extraction schemes tested on the same database and reported in [2] and [15]. The first method is applied to the lung area defined by [28] but not extended as we do (our lung area is about 1.5 their area), bringing to a loss of 12 true positives, out of 153, even before the candidate extraction. The result of the extraction scheme is a set of candidates and a loss of other 8 true positives; furthermore the authors apply a classification method that selects 5028 candidates, loosing other 15 true positives, for a total of 35 false negatives. We implemented the second method and applied it to same lung area used in [28] obtaining really poor results. To prune the set of extracted candidates we calculated for each region a set of 40 features and studied their distribution. The statistical analysis allowed us to select a set of 21 most representative features, based on the shape (10 features), the position (1 feature), the gray level distribution in the original image (2 features), and the values of the coefficients P R calculated during the candidates

6 extraction (7 features). Applying to these features simple thresholding rules and others describing the relationships observed between pairs of features, we can easily discard about false positives without loosing any true positive, hence reaching a sensitivity ratio equal to 0.96 and a total number of candidates equal to The high number of candidates obtained is mainly due to the fact that we use a lung area that is about 1.5 times bigger than the one commonly considered. 7 Future work The results obtained applying very simple rules prove the efficiency of the set of selected features and make us believe that using them for training suitable classifiers would improve the performances. Besides the system would be more general, hence more robust to the change of the database. References 1. Medical university of south carolina. Web Address: 2. B.Van Ginneken A. Schilham and M. Loog. Multi-scale nodule detection in chest radiographs. Proc. MICCAI, S.G. Armato, M.Giger, and H.MacMahon. Automated lung segmentation in digitized posteroanterior chest radiographs. Academic radiology, 5: , M.S. Brown, L.S. Wilson, B.D. Doust, R.W. Gill, and C.Sun. Knowledge-based method for segmentation and analysis of lung boundaries in chest x-rays images. Computerized Medical Imaging and Graphics, 22: , F.M. Carrascal, J.M. Carreira, M. Souto, P.G. Tahoces, L. Gomez, and J.J. Vidal. Automatic calculation of total lung capacity from automatically traced lung boundaries in postero- anterior and lateral digital chest radiographs. Medical Physics, 25: , J. Duryea and J.M. Boone. A fully automatic algorithmfor the segmentation of lung fields in digital chest radiographic images. Medical Physics, 22: , M. Giger, K. Doi, and H. Mac Mahon. Image feature analisys and computer-aided diagnosis in digital radiography:automated detection of nodules in peripheral lung fields. Med. Phisycs, 15: , M. Giger, K. Doi, H. Mac Mahon, C. Metz, and F.Yin. Pulmonary nodules: Computer-aided detection in digital chest images. Radiographics, 10:41 51, M. L. Giger, N. Ahn, K. Doi, H. MacMahon, and C. E. Metz. Computerized detection of pulmonary nodules in digital chest images: use of morphological filters in reducing false positive detections. Med. Phys., 17: , D.Catarious Jr., A.Baydush, and Jr. C.Floyd. Initial development of a computer aided diagnosis tool for solitary polmunary nodules. Proc SPIE, 4322: , J.Wei, Y. Hagihara, and H. Kobatake. Detection of cancerous tumors on chest x-ray images- candidates detection filtering and its evaluation. Presented at the Int. Conf. on Image. Proc (ICIP 99), J.Wei, Y. Hagihara, A. Shimizu, and H. Kobatake. Optimal image feature set for detecting lung nodules on chest x-rays images. CARS 2002, 2002.

7 13. A. Kano, K. Doi, H. MacMahon, D.D. Hassel, and M.Giger. Digital image subtraction of temporally sequential chest images for detection of interval change. Med. Physics, 21: , J. N. Kapur, P. K. Sahoo, and A.K. C.Woong. A new method for gray level picture thresholding using the entropy of the histogram. Computer Vision Graphics and Image Processing, 29: , Bilgin Keserci and Hiroyuki Yoshida. Computerized detection of pulmonary nodules in chest radiographs based on morphological features and wavelet snake model. Medical Image Analisys, 6: , J.-S. Lin, S.-C. Lo, M. Freedman, and S.Mun. Reduction of false positives in lung nodule detection using a two-level neural classification. IEEE Trans. Med. Imag., 15: , S.-C. Lo, M. Freedman, J.-S. Lin, and S.Mun. Automatic lung nodule detection using profile matching and backpropagation neural network techniques. Journal Digital Imaging, 1:48 54, S.-C. Lo, S.L.Lou, J.-S. Lin, M. Freedman, M.Chien, and S.Mun. Artificial convolution neural network techniques and applications for lung nodule detection. IEEE Trans. on Med. Imag., 14: , H. MacMahon and K. Doi. Digital chest radiography. Clin. Chest Med., 12:19 32, T. Matsumoto, H. Yoshimura, K. Doi, M. Giger, A. Kano, H. MacMahon, M. Abe, and S. Montner. Image feature analisys of false-positives diagnosis produced by automated detection of lung nodules. Investigative Radiol., 27: , M.F. McNitt-Gray, H.K. Huang, and J.W. Sayre. Feature selection in the pattern classification problem of digital chest radiographs segmentation. IEEE Trans. on Med. Imaging, 14: , M. Penedo, M. Carreira, A. Mosquera, and D. Cabello. Computer aided diagnosis: A neural network based approach to lung nodule detection. IEEE Trans. Med. Imag., 17: , J. Shiraishi, S. Katsuragawa, J. Ikezoe, T. Matsumoto, T. Kobayashi, K. Komatsu, M. Matsui, H. Fujita, Y. Kodera, and K. Doi. Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists detection of pulmonary nodules. AJR, 174:71 74, H. Suzuki and N. Inakoa. Development of a computer aided detection system for lung cancer diagnosis. Proc. SPIE, 1652: , H. Suzuki, N. Inakoa, H. Takabatake, M. Mori, H. Natori, and A. Suzuki. An experimental system for detecting lung nodules by chest x-ray image processing. Proc. SPIE, 1450:99 107, O. Tsuji, M.T. Freedman, and S.K. Mun. Automated segmentation of anatomic regions in chest radiographs using an adaptive-sized hybrid neural network. Med. Phys., 25: , B. van Ginneken. Computer-aided diagnosis in chest radiographs. P.h.D. dissertation, Utrecht Univ., Utrecht, The Nederlands, B. van Ginneken and B.M. ter H. Romeny. Automatic segmentation of lung fields in chest radiographs. Medical Physics, 27: , N.F. Vittitoe, R. Vargas-Voracek, and C.E. Floyd Jr. Markov random field modeling in posteroanterior chest radiograph segmentation. Med. Phys., 26: , Cj Vyborny. The aapm/rsna physics tutorial for residents: Image quality and the clinical radiographic examination. Radiographics, 17: , 1997.

8 31. Y. Wu, K. Doi, M. Giger, C. Metz, and W. Zhang. Reduction of false positives in computerized detection of lung nodules in chest radiographs using convolution neural network techniques and application for lung nodule detection. J. Digital Imaging, 7: , X.-W. Xu and K. Doi. Image features analysis for computer aided diagnosis:accurate determination of ribcage boundary in chest radiographs. Med.Phys., 22: , X.-W. Xu, S. Katsuragawa, A. Ashizawa, H. MacMahon, and K. Doi. Analisys of image features of histograms of edge gradient for false positive reduction in lung nodule detection in chest radiographs. Proc SPIE, 3338: , H. Yoshida, X. Xu, K. Doi, and M. Giger. Computer-aided diagnosis (cad) scheme for detecting pulmonary nodules using wavelet transforms. Proc. SPIE, 2434: , H. Yoshimura, M. Giger, K.Doi, H. MacMahon, and S. Montner. Computerized scheme for the detection of pulmonary nodules: A non linear filtering technique. Investigative radiol., 27: , 1992.