Ultrasound in Med. & Biol., Vol. -, No. -, pp. 1–8, 2014 Copyright 2014 World Federation for Ultrasound in Medicine & Biology Printed in the USA. All rights reserved 0301-5629/$ - see front matter
http://dx.doi.org/10.1016/j.ultrasmedbio.2014.09.004
d
Original Contribution VALIDATION OF A COMPUTER-AIDED DIAGNOSIS SYSTEM FOR THE AUTOMATIC IDENTIFICATION OF CAROTID ATHEROSCLEROSIS LILLA BONANNO,* SILVIA MARINO,*y PLACIDO BRAMANTI,* and FABRIZIO SOTTILEz * IRCCS Centro Neurolesi ‘‘Bonino-Pulejo’’, Messina, Italy; y Department of Biomedical Sciences and Morphologic and Functional Imaging, University of Messina, Messina, Italy; and z Department of Clinical and Experimental Medicine, University of Messina, Messina, Italy (Received 9 January 2014; revised 14 August 2014; in final form 2 September 2014)
Abstract—Carotid atherosclerosis represents one of the most important causes of brain stroke. The degree of carotid stenosis is, up to now, considered one of the most important features for determining the risk of brain stroke. Ultrasound (US) is a non-invasive, relatively inexpensive, portable technique, which has an excellent temporal resolution. Computer-aided diagnosis (CAD) has become one of the major research fields in medical and diagnostic imaging. We studied US images of 44 patients, 22 patients with and 22 without carotid artery stenosis, by using US examination and applying a CAD system, an automatic prototype software to detect carotid plaques. We obtained 287 regions: 60 were classified as plaques, with an average signal echogenicity of 244.1 ± 20.0 and 227 were classified as non-plaques, with an average signal echogenicity of 193.8 ± 38.6 compared with the opinion of an expert neurologist (golden test). The receiver operating characteristic (ROC) analysis revealed a highly significant area under the ROC curve difference from 0.5 (null hypothesis) in the discrimination between plaques and non-plaques; the diagnostic accuracy was 89% (95% CI: 0.85–0.92), with an appropriate cut-off value of 236.8, sensitivity was 83% and specificity reached a value of 85%. The experimental results showed that the proposed method is feasible and has a good agreement with the expert neurologist. Without the need of any user-interaction, this method generates a detection out-put that may be useful in second opinion. (E-mail: [email protected]) 2014 World Federation for Ultrasound in Medicine & Biology. Key Words: CAD system, Carotid atherosclerosis, Ultrasound image, Watershed segmentation.
of velocity imaging was originally based on the Doppler effect and is therefore often referred to as Doppler imaging (Cloutier et al. 2001). Computer-aided diagnosis (CAD) has become one of the major research directions in the medical imaging field (Doi 2007). CAD should be considered as an objective technique that aims to achieve both goals of lowering cost and effectiveness and is especially well suited for the digital imaging technology, which is being developed to produce digital images (Petrick et al. 1996). Recently, CAD is beginning to be applied widely in the detection and differential diagnosis of many different types of abnormalities in medical images (Doi et al. 1992; Summers 2003; Giger 2004; Doi 2005). Many different types of CAD systems are being developed for detection and/or characterization of lesions in medical imaging, including conventional radiography, computed tomography, magnetic resonance imaging and US (Doi et al. 1999, 2005). In literature, there are studies that present a CAD-based technique (atheromatic system) for classification of carotid plaques in B-mode US images
INTRODUCTION The degree of carotid stenosis is, up to now, considered one of the most important features for determining the risk of brain stroke (Afonso et al. 2012). This feature, together with other patient information such as age, clinical history and risk factors, are the main criteria for determining the risk of stroke and, thus, to decide about eventual surgical intervention (Nicolaides et al. 1995). Ultrasound (US) is a non-invasive relatively inexpensive, portable technique, which has an excellent temporal resolution (Suetens 2002). US imaging is used not only to visualize the anatomy of the stenosis but also to visualize function (Rangaraj 2005). Vascular non-invasive US allows the estimation of morphologic and dynamic parameters of arteries, such as diameter and distension or intima-media thickness (Rossi et al. 2008). The principle
Address correspondence to: Lilla Bonanno, IRCCS Centro Neurolesi ‘‘Bonino-Pulejo’’, S.S. 113, Via Palermo, Cntr. Casazza, 98124 Messina, Italy. E-mail: [email protected] 1
2
Ultrasound in Medicine and Biology
into symptomatic or asymptomatic classes (Acharya et al. 2012a, 2012b). The general approach for CAD system is to find the exact location of a lesion and also to determine an estimate of the probability of an atherosclerotic disease. The aim of this study was to develop a CAD system capable of discriminating the plaques and non-plaques and to identify the location and size of each plaque. We validated a CAD system in a set of US images by using ROC analysis (Zweig and Campbell 1993; Metz 2000). From the obtained results, we could argue that the system is accurate and clinicians could use the computer output as a ‘‘second opinion’’ to support the decision making. MATERIALS AND METHODS Study population We studied 44 patients, 22 with and 22 without carotid artery stenosis. All patients were randomly enrolled. The Watershed algorithm, implemented using MATLAB 7.6, was tested on US images. For all of the patients, the analysis was performed including the anamnestic risk clinical factors (diabetes, smoking, hypertension, dyslipidemia). The patients (mean age 63.82 6 16.66 y) presented stenosis at common (CCA), internal (ICA) and external (ECA) carotid artery between 20%–60% with about 35% median and have all risk factors that generate the formation of atherosclerotic plaque. The 22 patients without stenosis (mean age 57.04 6 21.04 y) presented very low risk factor levels. Detailed socio-demographic characteristics are summarized in Table 1. The patients were recruited from IRCCS Centro Neurolesi ’’BoninoPulejo’’ of Messina. Local Ethics Committee approval was obtained and all patients gave informed consent. Instruments and ultrasonography data acquisition The CCA, ICA and ECA US data were obtained as longitudinal cross-sections using a Philips iU22 ultraTable 1. Socio-demographic characteristics of patients Patients No. Patients Age Mean SD Gender M F Smoker (%) Ex-Smoker (%) Non-Smoker (%) Diabetes (%) Dyslipidemia (%) Hypertension (%)
Controls
22
22
63.8 16.7
57 21
10 12 27.3 31.8 40.9 22.7 72.7 59.1
4 18 4.5 9.1 86.4 1.3 2.5 40.9
Volume -, Number -, 2014
sound system (Philips Healthcare, Eindhoven, The Netherlands) with an L9-3 probe and included B-Mode (i.e., gray scale) and color Doppler image sequences. The vascular carotid preset on the machine was used (Vasc Car preset, persistence low, XRES and SONOCT on), and the gain was optimized by the operator (F.S.) who is an experienced vascular sonographer. We used 44 echo color Doppler images that were stored in a database to be read by the algorithm automatically and sequentially. Plaque segmentation algorithm The dynamic series were retrospectively transferred as anonymized Digital Imaging and Communications in Medicine (DICOM) files to the CAD system. The algorithm implemented a series of processing steps. After reading the B-mode image, a filter was applied (pre-processing phase) to obtain a better segmentation gradient. Watershed technique was used to segment the carotid plaques (processing phase). The B-mode features included average signal echogenicity on plaque region (features extraction phase), including finally, the classification phase of plaques. A similar previous prototype of the CAD system was described in detail by Mayer et al. (2006). The analysis of the images was automatically performed without any user interaction. A workflow diagram is shown in Figure 1. Pre-Processing The segmentation of echo color Doppler images is difficult because of variable imaging parameters, overlapping intensities, noise, gradients, motion, blurred edges and normal anatomic variation artifacts. The US artifacts can be classified as to their sources, which are physiologic (e.g., motion, different speeds of sound and acoustic impedance of tissues), equipment (dimension of the ultrasound beam and the converter array) and technical imaging (mode B, spectral Doppler and color Doppler ultrasound) (Amir et al., 2013). Therefore, before applying any approach to carotid artery stenosis, there are generally two pre-processing steps that have to be carried out first, the removal of artifacts from images and the removal of non-plaque features from the image. We considered 44 images, reporting the results obtained on a single image (Fig. 2) of a patient affected by a fibrocalcific atherosclerotic plaque of 25%–30% grade of stenosis localized in the bifurcation of the common carotid artery. As we wanted to maximize performance of the image segmentation methods, it was necessary to remove image inhomogeneities generated by the bias field and suppress the random noise generated by digital acquisition. This caused difficulties in applying techniques for the recovery of the contour of the object. We have applied a filter for removal of artifacts from images. In particular, we applied Sobel filter to detect edges of the image. The
Carotid atherosclerosis identification by CAD d L. BONANNO et al.
3
Fig. 3. Sobel filter image.
Fig. 1. Flowchart of the proposed CAD system. CAD 5 computer-aided diagnosis.
Sobel operator calculates the gradient of the image intensity at each point, giving the direction of the largest possible increase from light to dark and the rate of change in that direction. In Figure 3, the obtained image with filter application is shown. Processing To determine foreground objects in the image through the Watershed technique, we used the flooding algorithm (Vincent and Soille 1991). Let the original
Fig. 2. Ultrasound image of a 25%–30% grade of stenosis localized in the bifurcation common carotid artery.
gray-scale image be I, the gradient image VI is then computed. Image gradient is analogueous to the hills and hollows of a landscape, and the analoguey is continued by imagining rain pouring over the landscape, where the water falling on the landscape would flow down to the minimum. Thus, in an immersive simulation of this landscape, water floods from catchment basins when the altitude reaches the local maximum. A dam is therefore built to prevent the basins from merging when two floods originating from different catchment basins meet (Fig. 4). Direct application of Watershed transformation on the gradient images produce typically severe segmentation of the image because of numerous minima that are present in real (gradient) images due to inherent noise. One possibility to get rid of the false regions is the socalled ‘marker image’ used to mark those regions that require segmentation, although it is generally difficult to obtain relevant markers automatically without any interaction by the user (Parvati et al. 2008). A variety of procedures could be applied to find the foreground markers, which must connect blobs of pixels inside each of the foreground objects. The markers computation was done by using the morphologic operations called opening by reconstruction and closing by reconstruction
Fig. 4. Watershed algorithm diagram.
4
Ultrasound in Medicine and Biology
to clean up the image from stem and dark spots and removing the small blemishes without affecting the overall shape of the segmented objects. We used erosionbased gray-scale reconstruction (eqn [1]): \ ðrecÞ (1) fI ðJÞ 5 n$1 εn ðJÞ where εn ðJÞ can be obtained by iterating n elementary geodesic erosion, which is defined as (eqn [2]) ðIÞ
ε ðJÞ 5 ðJ.bÞWI
The parameters extracted from each ROI are those that are taken into account also by clinicians to describe the morphology of stenosis perimeter, area, distance, average signal echogenicity and centroid. For each ROI, we considered the average signal echogenicity feature. This parameter is normally expected to have values between 0 and 1 (non-plaque), (eqn [5]): 0#
Echogenicity Meansingle ROI #1 Echogenicity Meanimage
(5)
(2)
where b is the flat structuring element of size I and W stands for point wise maximum. Followed by dilation-based gray-scale reconstruction (eqn [3]): [ grec dn ðJÞ (3) I ðJÞ 5 n$1 where dI ðJÞ can be obtained by iterating an elementary geodesic dilation, which is defined as (eqn [4]) \ dðIÞ ðJÞ 5 ðJ4bÞ I (4) T where b is the flat structuring element of size I and stands for point wise minimum. These techniques are more effective at removing small blemishes without affecting the overall shapes of the objects. The computation of the regional maxima of these reconstructed images is done to get smooth edge foreground objects. Then, we computed background markers and applied transformed Watershed for segmentation. The obtained image after Watershed segmentation is a color image in which the region of interest (ROI) found is colored. Finally, we superimposed this image (pseudo-color label matrix) on top of the original image (Fig. 5). Successively, the algorithm applied the cluster analysis on some parameters extracted from the plaques. Feature extraction To describe morphologic characteristics, 12 shapebased and texture feature parameters were calculated.
Fig. 5. Image obtained with Watershed algorithm.
Volume -, Number -, 2014
If the output value exceeds 1, the system classified the region as plaques (eqn [6]): Echogenicity Meansingle ROI $1 Echogenicity Meanimage
(6)
The regions that satisfy the condition (eqn [6]) correspond to suspected regions, so we built a minimal set of three parameters: average signal echogenicity, distance of image center and centroid. We have calculated the distances between each ROI and the center of the image, and we calculated the centroids of each ROI. Classification We have applied the Ward’s Method (1963) norms to create clusters. Ward’s method uses a ‘‘merging cost’’ for combining pairs of clusters. The measure used is the increase in the total within-cluster sum of squares when the two clusters are merged (eqn [7]): X x i 2! m AWB k2 DðA; BÞ 5 k! i˛AWB
2
X i˛A
k! x i 2! m Ak 1 2
X
k! x i 2! m Bk
! (7) 2
i˛B
where ! m j is the center of cluster j, and kk is Euclidean distance. The left sum in DðA; BÞ is the total within-cluster sum of squares in the combined cluster, and the corresponding sums of squares in clusters A and B are subtracted from it. The sum of squares measure DðA; BÞ is equivalent to the following distance measure (eqn [8]): rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi nA nB ! ! 2 (8) dðA; BÞ 5 k m A2 m Bk nA 1nB where nj is the number of elements contained in cluster j. In the process of agglomeration, the matrix is created to classify different ROI and the most similar ones are joined together to form a cluster. Hierarchical cluster analysis (HCA) was applied to the parameters to solve the problem of undesirable over-segmentation results produced by the Watershed
Carotid atherosclerosis identification by CAD d L. BONANNO et al.
technique. HCA was placed inside the Watershed algorithm. HCA presented as dendrogram indicates the compatibility among different parameters (average signal echogenicity, distance of image center and centroid) and their mixtures from two separate clusters at the linkage distance of 10 (Fig. 6). We obtained two clusters, red and blue, which means that the ROI have similar characteristics to be grouped in a red or blue cluster. In particular, by matching the identification of the ROI, we can argue that the ROI of blue group are plaques. This deduction is confirmed by the clinical analysis of Figure 1, which identified this area as plaque. In Figure 7, the obtained image after cluster analysis is shown. Diagnostic accuracy ROC analysis was performed to calculate the diagnostic accuracy (area under the ROC curve, AUC) of the CAD system (null hypothesis: AUC 5 0.5), with an appropriate cut-off. In this case, we compared the obtained results on 44 echo color Doppler images containing plaque with the opinion of an expert neurologist (golden test). Sensitivity, specificity, negative and positive predictive values, positive and negative likelihood ratios at designated cut-off levels with their 95% confidence interval (CI) were evaluated. p value less than 0.05 (twosided) was considered to indicate statistical significances. RESULTS The proposed method was applied to 44 echo color Doppler images, 22 with and 22 without carotid artery stenosis. We obtained 287 regions, where 60 were classified as plaques, with an average signal echogenicity of 244.1 6 20.0, and 227 were classified as non-plaques, with an average signal echogenicity of 193.8 6 38.6, compared with the opinion of an expert neurologist (golden test). The analysis has been performed using
Fig. 6. Dendrogram result obtained by applying the clustering algorithm of Ward’s Method. The region of interest of blue cluster is identified as plaques.
5
Fig. 7. Image obtained after cluster analysis.
the software R 2.15. A hypothesis test of interest is whether the level of signal average echogenicity discriminates the two groups. This problem can be reformulated in terms of AUC. The level of echogenicity average was measured on a continuous scale with observations on independent groups. The AUC is equal to 0.89, value for which the marker would be moderately accurate, according to the classification of Swets (1988). We want to determine the threshold value, which discriminates between the two groups. The optimal threshold value will be the point with shorter distance (Perkins et al., 2007). In this case, the output of the statistical tests shows that, for the Youden index, the optimal cut-off value is k 5 236.8 (Table 2). The ROC analysis revealed a highly significant AUC difference from 0.5 (null hypothesis) in the discrimination between plaques and non-plaques; the diagnostic accuracy was 89% (95% CI: 0.85–0.92), with an appropriate cut-off value of 236.8, sensitivity was 83% and specificity reached a value of 85% (Fig. 8). DISCUSSION Quantitative characterization of carotid atherosclerosis and classification of plaques is crucial in the diagnosis and treatment planning. For this reason, there is a need for an automated system that permits the identification of plaques and differentiation, also to avoid the subjectivity of the clinician in identifying plaques. Recently, CAD systems have been developed to improve the capability of clinician interpretation of medical images and differentiation between benign and malignant tissues (Doi et al. 1999; Giger 2000; Vyborny et al. 2000; Giger and Karssemeijer 2001). The efficiency of clinicians’ interpretation can be improved in terms of accuracy and consistency in detection/diagnosis, while their productivity can be improved by reducing the time required for reading the images (Doi 2009). The computer outputs are derived using various techniques in computer vision to present some of the significant parameters such as the location of suspected lesions and the likelihood of malignancy of detected lesions. Generally,
6
Ultrasound in Medicine and Biology
Volume -, Number -, 2014
Table 2. Receiver operating characteristic analysis data Cut-off
TP
FP
TN
FN
Sens.
95% CI
Spec.
95% CI
PPV
NPV
AUC (95% CI)
.229.3 .231.7 .232.2 .236.8* .237.9 .238.1 .241
51 51 50 50 49 49 46
41 39 39 33 33 32 28
186 188 188 194 194 195 199
9 9 10 10 11 11 14
85.0 85.0 83.3 83.3 81.7 81.7 76.7
73.4–92.9 73.4–92.9 71.5–91.7 71.5–91.7 69.6–90.5 69.6–90.5 64–86.6
81.94 82.8 82.8 85.5 85.5 85.9 87.7
76.3–86.7 77.3–87.5 77.3–83.5 80.2–89.8 80.2–89.8 80.7–90.2 82.7–91.6
0.55 0.57 0.56 0.60 0.60 0.60 0.62
0.95 0.95 0.95 0.95 0.95 0.95 0.93
0.89 (0.85–0.92)
CI 5 confidence interval; TP 5 true positive; FP 5 false positive; TN 5 true negative; FN 5 false negative; Sens. 5 sensitivity; Spec. 5 specificity; PPV 5 positive predictive value; NPV 5 negative predictive value; AUC 5 area under the curve. * Criterion corresponding with highest Youden index.
CAD systems are executable on all imaging modalities and all kinds of examinations. The aim of this study was to develop a CAD system capable of discriminating the plaque from non-plaque features and also to identify the location and size of each plaque in the US images. In addition, cluster analysis was used to solve the problem of undersided, over-segmentation results produced by the Watershed technique and to reduce the number of false detections. From the obtained results by ROC curve, we can suppose that our algorithm could be particularly helpful for an objective identification of plaques. By the preliminary experimental results for 44 images, the proposed method can almost find all regions that are plaques, with a diagnostic accuracy of 89%. In fact, we have seen that echogenicity significantly differs between plaques and non-plaques. To improve specificity without significantly reducing sensitivity, morphologic features have been implemented in clinical routine as further diagnostic criteria in US image. Our algorithm uses the gray-scale image converted from the color image. These parameters had a translational effect related to the segmentation of plaques with different characteristics of shape, size and contrast. From ROC curve analysis, we have obtained that if the level of signal average echogenicity is more than 236.8 (cut-off k 5 236.8) the ROI is considered as a plaque, with a diagnostic accuracy of 89%. Then, the algorithm allows us to extract information about a ROI descriptive morphology if we observe the value of the average signal echogenicity features, we can demonstrate if it is plaque or not. In previous CAD studies, B-mode US images were used for classification of carotid plaques into symptomatic or asymptomatic classes (Acharya et al. 2012a). In this study, the authors used a combination of discrete wavelet transform and averaging algorithms. In addition, they used a support vector machine (SVM) classifier for automated decision making. Their results showed a classification accuracy of 82.4%, sensitivity of 82.9% and specificity of 82.1%. The authors showed that texture features coupled with the SVM classifier can be used to identify the plaque tissue type. Although
the number of ROI they get is greater than that selected by our algorithm, the accuracy (83.7%) is lower than the one obtained in our results (89%). This could be due to the fact that the segmentation is manually performed so it is difficult to exactly compare these results with our work, where segmentation is totally automatic. Acharya et al. (2013) recently presented a CAD system using two different databases consisting of images that characterize the textural differences in these symptomatic and asymptomatic classes using several grayscale features based on a novel combination of trace transform and fuzzy texture. In this case, the accuracy in first database is higher than ours. Seabra et al. (2011) developed an ultrasound-based diagnostic measure that quantifies plaque activity (the likelihood of the asymptomatic lesion to produce neurologic symptoms). This information is used to build an enhanced activity index, which considers the conditional probabilities of each relevant feature belonging to either symptomatic or asymptomatic groups. This measure was evaluated on a longitudinal study of 112 asymptomatic plaques and shows high diagnostic power. Also, in this case the processing step consists of a manual segmentation. Other works treat the most recent advances in breast cancer detection/diagnosis using CAD systems developed for mammography and ultrasound examination.
Fig. 8. The considered receiver operating characteristic curve for average signal echogenicity. The value of the classification result produced best performance with an area under the curve value of 0.89.
Carotid atherosclerosis identification by CAD d L. BONANNO et al.
Lai et al. (2013) proposed a CAD system to quantify the tumor morphology of vascularity on 3-D power Doppler breast ultrasound images. However, the accuracy we showed is higher than that reported by Lai et al. (2013), even if their results showed that the combination of the B-mode and vascular features has improved performance compared with single feature types of B-mode or vascularity. Comparing the results obtained, we can argue that the CAD system is a very good segmentation algorithm for US images. The limitation of this study is that the algorithm is not able to identify the consistency of plaque and fails to define the morphology. In addition, the measured video intensity cannot be easily translated to the local reflection coefficient, the physical parameter that probably differentiates atherosclerotic plaque from other tissues. This may increase the standard deviation of the ‘‘average signal echogenicity’’ for regions with and without stenosis, and reduce the technique’s sensitivity and specificity. This limitation may be overcome by adaptively setting the video intensity threshold for each region of interest based on statistical analysis of the gray levels in the surrounding area, defined in such a way that all operator dependent parameters, as well as the cumulative attenuation between the probe and each pixel within the area, are similar. Thus, the results obtained show a high sensitivity and specificity in comparison with other studies in the literature but they are considered only relatively about the parameters obtained by US examination. Unlike studies in literature, we have generated a totally automated system, and we have created a minimal set of parameter descriptors ROI. In particular, we have found that the echogenicity may be a good marker to discriminate plaques from non-plaques. CONCLUSIONS Doppler ultrasound is an operator-dependent modality, and the interpretation of the obtained images requires a very high expertise. To overcome the operator dependency and increase accurate diagnosis rate, CAD systems are developed for detection and classification of plaques. For this reason there is a need for an automated system that permits the identification and differentiation of plaques also to avoid the subjectivity of the clinician in the correct morphologic identification and localization of plaques. To date, it is the first study that reported a fully automatic CAD system used to distinguish between plaque and non-plaque areas. The computer extracted average echogenicity features as well as the final classification significantly differed between plaques and non-plaques. An observer independent CAD system may be a promising supplementary tool for interpreting carotid stenosis in US images.
7
Further investigation with larger and more heterogeneous patient samples is needed to consolidate this software prototype as a solid second opinion tool. REFERENCES Acharya UR, Faust O, Alvin APC, Sree SV, Molinari F, Saba L, Nicolaides A, Suri JS. Symptomatic versus asymptomatic plaque classification in carotid ultrasound. J Med Syst 2012a;3:1861–1871. Acharya UR, Molinari F, Saba L, Nicolaides A, Shafique S, Suri JS. Carotid ultrasound symptomatology using atherosclerosis plaque characterization: a class of atheromatic systems. Conf Proc IEEE Eng Med Biol Soc 2012b;2012:3199–3202. Acharya UR, Mookiah MR, Vinitha Sree S, Afonso D, Sanches J, Shafique S. Atherosclerotic plaque tissue characterization in 2-D ultrasound longitudinal carotid scans for automated classification: a paradigm for stroke risk assessment. Med Biol Eng Comput 2013; 51:513–523. Afonso D, Seabra J, Suri JS, Sanches JM. A CAD system for the atherosclerotic plaque assessment. 34th Annual International Conference of the IEEE EMBS. San Diego: California; 2012. Amir S, Chowdhry BS, Hashmani M, Hasan M. The analysis of the artifacts due to the simultaneous use of two ultrasound probes with different/similar operating frequencies. Comput Math Methods Med 2013;2013:890170. Cloutier G, Chen D, Durand LG. Performance of time-frequency representation techniques to measure blood flow turbulence with pulsed wave Doppler ultrasound. Ultrasound Med Biol 2001;4:535–550. Doi K, Giger ML, MacMahon H, Hoffmann KR, Nishikawa RM, Schmidt RA, Chua KG, Katsuragawa S, Nakamori N, Sanada S, et al. Computer-aided diagnosis: development of automated schemes for quantitative analysis of radiographic images. Semin Ultrasound CT MR 1992;13:140–152. Doi K. Computer-aided diagnosis in medical imaging: Historical review, current status and future potential. Comput Med Imaging Graph 2007;31:198–211. Doi K. Current status and future potential of computer-aided diagnosis in medical imaging. The Br J Radiol 2005;78:S3–S19. Doi K, MacMahon H, Katsuragawa S, Nishikawa RM, Jiang Y. Computer-aided diagnosis in radiology potential and pitfalls. Eur J Radiol 1999;2:97–109. Doi K. Computer-aided diagnosis in medical imaging achievements and challenges. In: World Congress on Medical Physics and Biomedical Engineering. Munich, Germany: 11th International Congress of the IUPESM; 2009. Giger ML. Computer-aided diagnosis of breast lesions in medical images. Comput Sci Eng 2000;2:39–45. Giger ML, Karssemeijer N. Computer-aided diagnosis in medical imaging. IEEE Trans Med Imaging 2001;20:1205–1208. Giger ML. Computerized analysis of images in the detection and diagnosis of breast cancer Seminars in Ultrasound. CT MRI 2004;25: 411–418. Lai YC, Huang YS, Wang DW, Tiu CM, Chou YH, Chang RF. Computer-aided diagnosis for 3-D power Doppler breast ultrasound. Ultrasound Med Biol 2013;39:555–567. Mayer D, Vomweg TW, Faber H, Weinheimer O, Mattiuzzi M, Buscema M, D€uber C. Fully automatic breast cancer diagnosis in contrast enhanced MRI. Int J Comp Assist Radiol Surg 2006;1: 325–343. Metz CE. Fundamental ROC analysis. In: Beutel J, Kundel HL, Van Metter RL, (eds). Handbook of medical imaging. Bellingham, Washington: SPIE Press, The International Society for Optical Engineering; 2000. p. 751–770. Nicolaides AN, Barnett JHM, Belcaro GV, Bernstein EF, Callow A, Eastcott HHG, Eikelboom B, Eklof B, Geroulakos G, Halliday A, Hobson R, Kalodiki E, Lusby R, Malikides A, Middleton LT, Moore WS, Norris J, Prescott L, Ramaswami G, Rutherford R, Sandmann W, Shanik G, Spartera C, Thomas D. Consensus statement on the management of patients with asymptomatic atherosclerotic carotid bifurcation lesions. Int Angiol 1995;14:5–17.
8
Ultrasound in Medicine and Biology
Parvati K, Prakasa Rao BS, Mariya Das M. Image segmentation using gray-scale morphology and marker-controlled Watershed transformation. Hindawi Publishing Corporation. Discrete Dynamics in Nature and Society 2008;2008:1–8. Perkins NJ, Schisterman EF, Vexler A. ROC curve inference from a sample with a limit of detection. Am J Epidemiol 2007;165:325–333. Petrick N, Chan HP, Wei D, Sahiner B, Helvie MA, Adler DD. Automated detection of breast masses on mammograms using adaptive contrast enhancement and texture classification. Med Phys 1996; 23:1685–1696. Rangaraj MR. Biomedical image analysis. CRC Press; 2005. Rossi AC, Brands PJ, Hoeks APG. Automatic recognition of the common carotid artery in longitudinal ultrasound B-mode scans. Med Image Analysis 2008;12:653–665. Seabra J, Pedro ML, Fernandes JF, Sanches J. Ultrasound Plaque Enhanced Activity Index for Predicting Neurologic Symptoms.
Volume -, Number -, 2014 Pattern Recognition and Image Analysis Lecture Notes in Computer Science 2011;2011(6669):184–191. Suetens P. Fundamentals of medical imaging. Cambridge University Press; 2002. Summers RM. Road maps for advancement of radiologic computeraided detection in the 21st century. Radiology 2003;229:11–13. Swets JA. Measuring the accuracy of diagnostic systems. Science 1988; 4857:1285–1293. Vincent L, Soille P. Watersheds in digital spaces: an efficient algorithm based on immersion simulations. IEEE Trans Pattern Anal Mach Intel 1991;6:583–598. Vyborny CJ, Giger ML, Nishikawa RM. Computer-aided detection and diagnosis of breast cancer. Radiol Clin North Am 2000;38: 725–740. Zweig MH, Campbell G. Receiver-Operating Characteristic (ROC) plots a fundamental evaluation tool in clinical medicine. Clin Chem 1993;19:561–577.
http://dx.doi.org/10.1016/j.ultrasmedbio.2014.09.004
d
Original Contribution VALIDATION OF A COMPUTER-AIDED DIAGNOSIS SYSTEM FOR THE AUTOMATIC IDENTIFICATION OF CAROTID ATHEROSCLEROSIS LILLA BONANNO,* SILVIA MARINO,*y PLACIDO BRAMANTI,* and FABRIZIO SOTTILEz * IRCCS Centro Neurolesi ‘‘Bonino-Pulejo’’, Messina, Italy; y Department of Biomedical Sciences and Morphologic and Functional Imaging, University of Messina, Messina, Italy; and z Department of Clinical and Experimental Medicine, University of Messina, Messina, Italy (Received 9 January 2014; revised 14 August 2014; in final form 2 September 2014)
Abstract—Carotid atherosclerosis represents one of the most important causes of brain stroke. The degree of carotid stenosis is, up to now, considered one of the most important features for determining the risk of brain stroke. Ultrasound (US) is a non-invasive, relatively inexpensive, portable technique, which has an excellent temporal resolution. Computer-aided diagnosis (CAD) has become one of the major research fields in medical and diagnostic imaging. We studied US images of 44 patients, 22 patients with and 22 without carotid artery stenosis, by using US examination and applying a CAD system, an automatic prototype software to detect carotid plaques. We obtained 287 regions: 60 were classified as plaques, with an average signal echogenicity of 244.1 ± 20.0 and 227 were classified as non-plaques, with an average signal echogenicity of 193.8 ± 38.6 compared with the opinion of an expert neurologist (golden test). The receiver operating characteristic (ROC) analysis revealed a highly significant area under the ROC curve difference from 0.5 (null hypothesis) in the discrimination between plaques and non-plaques; the diagnostic accuracy was 89% (95% CI: 0.85–0.92), with an appropriate cut-off value of 236.8, sensitivity was 83% and specificity reached a value of 85%. The experimental results showed that the proposed method is feasible and has a good agreement with the expert neurologist. Without the need of any user-interaction, this method generates a detection out-put that may be useful in second opinion. (E-mail: [email protected]) 2014 World Federation for Ultrasound in Medicine & Biology. Key Words: CAD system, Carotid atherosclerosis, Ultrasound image, Watershed segmentation.
of velocity imaging was originally based on the Doppler effect and is therefore often referred to as Doppler imaging (Cloutier et al. 2001). Computer-aided diagnosis (CAD) has become one of the major research directions in the medical imaging field (Doi 2007). CAD should be considered as an objective technique that aims to achieve both goals of lowering cost and effectiveness and is especially well suited for the digital imaging technology, which is being developed to produce digital images (Petrick et al. 1996). Recently, CAD is beginning to be applied widely in the detection and differential diagnosis of many different types of abnormalities in medical images (Doi et al. 1992; Summers 2003; Giger 2004; Doi 2005). Many different types of CAD systems are being developed for detection and/or characterization of lesions in medical imaging, including conventional radiography, computed tomography, magnetic resonance imaging and US (Doi et al. 1999, 2005). In literature, there are studies that present a CAD-based technique (atheromatic system) for classification of carotid plaques in B-mode US images
INTRODUCTION The degree of carotid stenosis is, up to now, considered one of the most important features for determining the risk of brain stroke (Afonso et al. 2012). This feature, together with other patient information such as age, clinical history and risk factors, are the main criteria for determining the risk of stroke and, thus, to decide about eventual surgical intervention (Nicolaides et al. 1995). Ultrasound (US) is a non-invasive relatively inexpensive, portable technique, which has an excellent temporal resolution (Suetens 2002). US imaging is used not only to visualize the anatomy of the stenosis but also to visualize function (Rangaraj 2005). Vascular non-invasive US allows the estimation of morphologic and dynamic parameters of arteries, such as diameter and distension or intima-media thickness (Rossi et al. 2008). The principle
Address correspondence to: Lilla Bonanno, IRCCS Centro Neurolesi ‘‘Bonino-Pulejo’’, S.S. 113, Via Palermo, Cntr. Casazza, 98124 Messina, Italy. E-mail: [email protected] 1
2
Ultrasound in Medicine and Biology
into symptomatic or asymptomatic classes (Acharya et al. 2012a, 2012b). The general approach for CAD system is to find the exact location of a lesion and also to determine an estimate of the probability of an atherosclerotic disease. The aim of this study was to develop a CAD system capable of discriminating the plaques and non-plaques and to identify the location and size of each plaque. We validated a CAD system in a set of US images by using ROC analysis (Zweig and Campbell 1993; Metz 2000). From the obtained results, we could argue that the system is accurate and clinicians could use the computer output as a ‘‘second opinion’’ to support the decision making. MATERIALS AND METHODS Study population We studied 44 patients, 22 with and 22 without carotid artery stenosis. All patients were randomly enrolled. The Watershed algorithm, implemented using MATLAB 7.6, was tested on US images. For all of the patients, the analysis was performed including the anamnestic risk clinical factors (diabetes, smoking, hypertension, dyslipidemia). The patients (mean age 63.82 6 16.66 y) presented stenosis at common (CCA), internal (ICA) and external (ECA) carotid artery between 20%–60% with about 35% median and have all risk factors that generate the formation of atherosclerotic plaque. The 22 patients without stenosis (mean age 57.04 6 21.04 y) presented very low risk factor levels. Detailed socio-demographic characteristics are summarized in Table 1. The patients were recruited from IRCCS Centro Neurolesi ’’BoninoPulejo’’ of Messina. Local Ethics Committee approval was obtained and all patients gave informed consent. Instruments and ultrasonography data acquisition The CCA, ICA and ECA US data were obtained as longitudinal cross-sections using a Philips iU22 ultraTable 1. Socio-demographic characteristics of patients Patients No. Patients Age Mean SD Gender M F Smoker (%) Ex-Smoker (%) Non-Smoker (%) Diabetes (%) Dyslipidemia (%) Hypertension (%)
Controls
22
22
63.8 16.7
57 21
10 12 27.3 31.8 40.9 22.7 72.7 59.1
4 18 4.5 9.1 86.4 1.3 2.5 40.9
Volume -, Number -, 2014
sound system (Philips Healthcare, Eindhoven, The Netherlands) with an L9-3 probe and included B-Mode (i.e., gray scale) and color Doppler image sequences. The vascular carotid preset on the machine was used (Vasc Car preset, persistence low, XRES and SONOCT on), and the gain was optimized by the operator (F.S.) who is an experienced vascular sonographer. We used 44 echo color Doppler images that were stored in a database to be read by the algorithm automatically and sequentially. Plaque segmentation algorithm The dynamic series were retrospectively transferred as anonymized Digital Imaging and Communications in Medicine (DICOM) files to the CAD system. The algorithm implemented a series of processing steps. After reading the B-mode image, a filter was applied (pre-processing phase) to obtain a better segmentation gradient. Watershed technique was used to segment the carotid plaques (processing phase). The B-mode features included average signal echogenicity on plaque region (features extraction phase), including finally, the classification phase of plaques. A similar previous prototype of the CAD system was described in detail by Mayer et al. (2006). The analysis of the images was automatically performed without any user interaction. A workflow diagram is shown in Figure 1. Pre-Processing The segmentation of echo color Doppler images is difficult because of variable imaging parameters, overlapping intensities, noise, gradients, motion, blurred edges and normal anatomic variation artifacts. The US artifacts can be classified as to their sources, which are physiologic (e.g., motion, different speeds of sound and acoustic impedance of tissues), equipment (dimension of the ultrasound beam and the converter array) and technical imaging (mode B, spectral Doppler and color Doppler ultrasound) (Amir et al., 2013). Therefore, before applying any approach to carotid artery stenosis, there are generally two pre-processing steps that have to be carried out first, the removal of artifacts from images and the removal of non-plaque features from the image. We considered 44 images, reporting the results obtained on a single image (Fig. 2) of a patient affected by a fibrocalcific atherosclerotic plaque of 25%–30% grade of stenosis localized in the bifurcation of the common carotid artery. As we wanted to maximize performance of the image segmentation methods, it was necessary to remove image inhomogeneities generated by the bias field and suppress the random noise generated by digital acquisition. This caused difficulties in applying techniques for the recovery of the contour of the object. We have applied a filter for removal of artifacts from images. In particular, we applied Sobel filter to detect edges of the image. The
Carotid atherosclerosis identification by CAD d L. BONANNO et al.
3
Fig. 3. Sobel filter image.
Fig. 1. Flowchart of the proposed CAD system. CAD 5 computer-aided diagnosis.
Sobel operator calculates the gradient of the image intensity at each point, giving the direction of the largest possible increase from light to dark and the rate of change in that direction. In Figure 3, the obtained image with filter application is shown. Processing To determine foreground objects in the image through the Watershed technique, we used the flooding algorithm (Vincent and Soille 1991). Let the original
Fig. 2. Ultrasound image of a 25%–30% grade of stenosis localized in the bifurcation common carotid artery.
gray-scale image be I, the gradient image VI is then computed. Image gradient is analogueous to the hills and hollows of a landscape, and the analoguey is continued by imagining rain pouring over the landscape, where the water falling on the landscape would flow down to the minimum. Thus, in an immersive simulation of this landscape, water floods from catchment basins when the altitude reaches the local maximum. A dam is therefore built to prevent the basins from merging when two floods originating from different catchment basins meet (Fig. 4). Direct application of Watershed transformation on the gradient images produce typically severe segmentation of the image because of numerous minima that are present in real (gradient) images due to inherent noise. One possibility to get rid of the false regions is the socalled ‘marker image’ used to mark those regions that require segmentation, although it is generally difficult to obtain relevant markers automatically without any interaction by the user (Parvati et al. 2008). A variety of procedures could be applied to find the foreground markers, which must connect blobs of pixels inside each of the foreground objects. The markers computation was done by using the morphologic operations called opening by reconstruction and closing by reconstruction
Fig. 4. Watershed algorithm diagram.
4
Ultrasound in Medicine and Biology
to clean up the image from stem and dark spots and removing the small blemishes without affecting the overall shape of the segmented objects. We used erosionbased gray-scale reconstruction (eqn [1]): \ ðrecÞ (1) fI ðJÞ 5 n$1 εn ðJÞ where εn ðJÞ can be obtained by iterating n elementary geodesic erosion, which is defined as (eqn [2]) ðIÞ
ε ðJÞ 5 ðJ.bÞWI
The parameters extracted from each ROI are those that are taken into account also by clinicians to describe the morphology of stenosis perimeter, area, distance, average signal echogenicity and centroid. For each ROI, we considered the average signal echogenicity feature. This parameter is normally expected to have values between 0 and 1 (non-plaque), (eqn [5]): 0#
Echogenicity Meansingle ROI #1 Echogenicity Meanimage
(5)
(2)
where b is the flat structuring element of size I and W stands for point wise maximum. Followed by dilation-based gray-scale reconstruction (eqn [3]): [ grec dn ðJÞ (3) I ðJÞ 5 n$1 where dI ðJÞ can be obtained by iterating an elementary geodesic dilation, which is defined as (eqn [4]) \ dðIÞ ðJÞ 5 ðJ4bÞ I (4) T where b is the flat structuring element of size I and stands for point wise minimum. These techniques are more effective at removing small blemishes without affecting the overall shapes of the objects. The computation of the regional maxima of these reconstructed images is done to get smooth edge foreground objects. Then, we computed background markers and applied transformed Watershed for segmentation. The obtained image after Watershed segmentation is a color image in which the region of interest (ROI) found is colored. Finally, we superimposed this image (pseudo-color label matrix) on top of the original image (Fig. 5). Successively, the algorithm applied the cluster analysis on some parameters extracted from the plaques. Feature extraction To describe morphologic characteristics, 12 shapebased and texture feature parameters were calculated.
Fig. 5. Image obtained with Watershed algorithm.
Volume -, Number -, 2014
If the output value exceeds 1, the system classified the region as plaques (eqn [6]): Echogenicity Meansingle ROI $1 Echogenicity Meanimage
(6)
The regions that satisfy the condition (eqn [6]) correspond to suspected regions, so we built a minimal set of three parameters: average signal echogenicity, distance of image center and centroid. We have calculated the distances between each ROI and the center of the image, and we calculated the centroids of each ROI. Classification We have applied the Ward’s Method (1963) norms to create clusters. Ward’s method uses a ‘‘merging cost’’ for combining pairs of clusters. The measure used is the increase in the total within-cluster sum of squares when the two clusters are merged (eqn [7]): X x i 2! m AWB k2 DðA; BÞ 5 k! i˛AWB
2
X i˛A
k! x i 2! m Ak 1 2
X
k! x i 2! m Bk
! (7) 2
i˛B
where ! m j is the center of cluster j, and kk is Euclidean distance. The left sum in DðA; BÞ is the total within-cluster sum of squares in the combined cluster, and the corresponding sums of squares in clusters A and B are subtracted from it. The sum of squares measure DðA; BÞ is equivalent to the following distance measure (eqn [8]): rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi nA nB ! ! 2 (8) dðA; BÞ 5 k m A2 m Bk nA 1nB where nj is the number of elements contained in cluster j. In the process of agglomeration, the matrix is created to classify different ROI and the most similar ones are joined together to form a cluster. Hierarchical cluster analysis (HCA) was applied to the parameters to solve the problem of undesirable over-segmentation results produced by the Watershed
Carotid atherosclerosis identification by CAD d L. BONANNO et al.
technique. HCA was placed inside the Watershed algorithm. HCA presented as dendrogram indicates the compatibility among different parameters (average signal echogenicity, distance of image center and centroid) and their mixtures from two separate clusters at the linkage distance of 10 (Fig. 6). We obtained two clusters, red and blue, which means that the ROI have similar characteristics to be grouped in a red or blue cluster. In particular, by matching the identification of the ROI, we can argue that the ROI of blue group are plaques. This deduction is confirmed by the clinical analysis of Figure 1, which identified this area as plaque. In Figure 7, the obtained image after cluster analysis is shown. Diagnostic accuracy ROC analysis was performed to calculate the diagnostic accuracy (area under the ROC curve, AUC) of the CAD system (null hypothesis: AUC 5 0.5), with an appropriate cut-off. In this case, we compared the obtained results on 44 echo color Doppler images containing plaque with the opinion of an expert neurologist (golden test). Sensitivity, specificity, negative and positive predictive values, positive and negative likelihood ratios at designated cut-off levels with their 95% confidence interval (CI) were evaluated. p value less than 0.05 (twosided) was considered to indicate statistical significances. RESULTS The proposed method was applied to 44 echo color Doppler images, 22 with and 22 without carotid artery stenosis. We obtained 287 regions, where 60 were classified as plaques, with an average signal echogenicity of 244.1 6 20.0, and 227 were classified as non-plaques, with an average signal echogenicity of 193.8 6 38.6, compared with the opinion of an expert neurologist (golden test). The analysis has been performed using
Fig. 6. Dendrogram result obtained by applying the clustering algorithm of Ward’s Method. The region of interest of blue cluster is identified as plaques.
5
Fig. 7. Image obtained after cluster analysis.
the software R 2.15. A hypothesis test of interest is whether the level of signal average echogenicity discriminates the two groups. This problem can be reformulated in terms of AUC. The level of echogenicity average was measured on a continuous scale with observations on independent groups. The AUC is equal to 0.89, value for which the marker would be moderately accurate, according to the classification of Swets (1988). We want to determine the threshold value, which discriminates between the two groups. The optimal threshold value will be the point with shorter distance (Perkins et al., 2007). In this case, the output of the statistical tests shows that, for the Youden index, the optimal cut-off value is k 5 236.8 (Table 2). The ROC analysis revealed a highly significant AUC difference from 0.5 (null hypothesis) in the discrimination between plaques and non-plaques; the diagnostic accuracy was 89% (95% CI: 0.85–0.92), with an appropriate cut-off value of 236.8, sensitivity was 83% and specificity reached a value of 85% (Fig. 8). DISCUSSION Quantitative characterization of carotid atherosclerosis and classification of plaques is crucial in the diagnosis and treatment planning. For this reason, there is a need for an automated system that permits the identification of plaques and differentiation, also to avoid the subjectivity of the clinician in identifying plaques. Recently, CAD systems have been developed to improve the capability of clinician interpretation of medical images and differentiation between benign and malignant tissues (Doi et al. 1999; Giger 2000; Vyborny et al. 2000; Giger and Karssemeijer 2001). The efficiency of clinicians’ interpretation can be improved in terms of accuracy and consistency in detection/diagnosis, while their productivity can be improved by reducing the time required for reading the images (Doi 2009). The computer outputs are derived using various techniques in computer vision to present some of the significant parameters such as the location of suspected lesions and the likelihood of malignancy of detected lesions. Generally,
6
Ultrasound in Medicine and Biology
Volume -, Number -, 2014
Table 2. Receiver operating characteristic analysis data Cut-off
TP
FP
TN
FN
Sens.
95% CI
Spec.
95% CI
PPV
NPV
AUC (95% CI)
.229.3 .231.7 .232.2 .236.8* .237.9 .238.1 .241
51 51 50 50 49 49 46
41 39 39 33 33 32 28
186 188 188 194 194 195 199
9 9 10 10 11 11 14
85.0 85.0 83.3 83.3 81.7 81.7 76.7
73.4–92.9 73.4–92.9 71.5–91.7 71.5–91.7 69.6–90.5 69.6–90.5 64–86.6
81.94 82.8 82.8 85.5 85.5 85.9 87.7
76.3–86.7 77.3–87.5 77.3–83.5 80.2–89.8 80.2–89.8 80.7–90.2 82.7–91.6
0.55 0.57 0.56 0.60 0.60 0.60 0.62
0.95 0.95 0.95 0.95 0.95 0.95 0.93
0.89 (0.85–0.92)
CI 5 confidence interval; TP 5 true positive; FP 5 false positive; TN 5 true negative; FN 5 false negative; Sens. 5 sensitivity; Spec. 5 specificity; PPV 5 positive predictive value; NPV 5 negative predictive value; AUC 5 area under the curve. * Criterion corresponding with highest Youden index.
CAD systems are executable on all imaging modalities and all kinds of examinations. The aim of this study was to develop a CAD system capable of discriminating the plaque from non-plaque features and also to identify the location and size of each plaque in the US images. In addition, cluster analysis was used to solve the problem of undersided, over-segmentation results produced by the Watershed technique and to reduce the number of false detections. From the obtained results by ROC curve, we can suppose that our algorithm could be particularly helpful for an objective identification of plaques. By the preliminary experimental results for 44 images, the proposed method can almost find all regions that are plaques, with a diagnostic accuracy of 89%. In fact, we have seen that echogenicity significantly differs between plaques and non-plaques. To improve specificity without significantly reducing sensitivity, morphologic features have been implemented in clinical routine as further diagnostic criteria in US image. Our algorithm uses the gray-scale image converted from the color image. These parameters had a translational effect related to the segmentation of plaques with different characteristics of shape, size and contrast. From ROC curve analysis, we have obtained that if the level of signal average echogenicity is more than 236.8 (cut-off k 5 236.8) the ROI is considered as a plaque, with a diagnostic accuracy of 89%. Then, the algorithm allows us to extract information about a ROI descriptive morphology if we observe the value of the average signal echogenicity features, we can demonstrate if it is plaque or not. In previous CAD studies, B-mode US images were used for classification of carotid plaques into symptomatic or asymptomatic classes (Acharya et al. 2012a). In this study, the authors used a combination of discrete wavelet transform and averaging algorithms. In addition, they used a support vector machine (SVM) classifier for automated decision making. Their results showed a classification accuracy of 82.4%, sensitivity of 82.9% and specificity of 82.1%. The authors showed that texture features coupled with the SVM classifier can be used to identify the plaque tissue type. Although
the number of ROI they get is greater than that selected by our algorithm, the accuracy (83.7%) is lower than the one obtained in our results (89%). This could be due to the fact that the segmentation is manually performed so it is difficult to exactly compare these results with our work, where segmentation is totally automatic. Acharya et al. (2013) recently presented a CAD system using two different databases consisting of images that characterize the textural differences in these symptomatic and asymptomatic classes using several grayscale features based on a novel combination of trace transform and fuzzy texture. In this case, the accuracy in first database is higher than ours. Seabra et al. (2011) developed an ultrasound-based diagnostic measure that quantifies plaque activity (the likelihood of the asymptomatic lesion to produce neurologic symptoms). This information is used to build an enhanced activity index, which considers the conditional probabilities of each relevant feature belonging to either symptomatic or asymptomatic groups. This measure was evaluated on a longitudinal study of 112 asymptomatic plaques and shows high diagnostic power. Also, in this case the processing step consists of a manual segmentation. Other works treat the most recent advances in breast cancer detection/diagnosis using CAD systems developed for mammography and ultrasound examination.
Fig. 8. The considered receiver operating characteristic curve for average signal echogenicity. The value of the classification result produced best performance with an area under the curve value of 0.89.
Carotid atherosclerosis identification by CAD d L. BONANNO et al.
Lai et al. (2013) proposed a CAD system to quantify the tumor morphology of vascularity on 3-D power Doppler breast ultrasound images. However, the accuracy we showed is higher than that reported by Lai et al. (2013), even if their results showed that the combination of the B-mode and vascular features has improved performance compared with single feature types of B-mode or vascularity. Comparing the results obtained, we can argue that the CAD system is a very good segmentation algorithm for US images. The limitation of this study is that the algorithm is not able to identify the consistency of plaque and fails to define the morphology. In addition, the measured video intensity cannot be easily translated to the local reflection coefficient, the physical parameter that probably differentiates atherosclerotic plaque from other tissues. This may increase the standard deviation of the ‘‘average signal echogenicity’’ for regions with and without stenosis, and reduce the technique’s sensitivity and specificity. This limitation may be overcome by adaptively setting the video intensity threshold for each region of interest based on statistical analysis of the gray levels in the surrounding area, defined in such a way that all operator dependent parameters, as well as the cumulative attenuation between the probe and each pixel within the area, are similar. Thus, the results obtained show a high sensitivity and specificity in comparison with other studies in the literature but they are considered only relatively about the parameters obtained by US examination. Unlike studies in literature, we have generated a totally automated system, and we have created a minimal set of parameter descriptors ROI. In particular, we have found that the echogenicity may be a good marker to discriminate plaques from non-plaques. CONCLUSIONS Doppler ultrasound is an operator-dependent modality, and the interpretation of the obtained images requires a very high expertise. To overcome the operator dependency and increase accurate diagnosis rate, CAD systems are developed for detection and classification of plaques. For this reason there is a need for an automated system that permits the identification and differentiation of plaques also to avoid the subjectivity of the clinician in the correct morphologic identification and localization of plaques. To date, it is the first study that reported a fully automatic CAD system used to distinguish between plaque and non-plaque areas. The computer extracted average echogenicity features as well as the final classification significantly differed between plaques and non-plaques. An observer independent CAD system may be a promising supplementary tool for interpreting carotid stenosis in US images.
7
Further investigation with larger and more heterogeneous patient samples is needed to consolidate this software prototype as a solid second opinion tool. REFERENCES Acharya UR, Faust O, Alvin APC, Sree SV, Molinari F, Saba L, Nicolaides A, Suri JS. Symptomatic versus asymptomatic plaque classification in carotid ultrasound. J Med Syst 2012a;3:1861–1871. Acharya UR, Molinari F, Saba L, Nicolaides A, Shafique S, Suri JS. Carotid ultrasound symptomatology using atherosclerosis plaque characterization: a class of atheromatic systems. Conf Proc IEEE Eng Med Biol Soc 2012b;2012:3199–3202. Acharya UR, Mookiah MR, Vinitha Sree S, Afonso D, Sanches J, Shafique S. Atherosclerotic plaque tissue characterization in 2-D ultrasound longitudinal carotid scans for automated classification: a paradigm for stroke risk assessment. Med Biol Eng Comput 2013; 51:513–523. Afonso D, Seabra J, Suri JS, Sanches JM. A CAD system for the atherosclerotic plaque assessment. 34th Annual International Conference of the IEEE EMBS. San Diego: California; 2012. Amir S, Chowdhry BS, Hashmani M, Hasan M. The analysis of the artifacts due to the simultaneous use of two ultrasound probes with different/similar operating frequencies. Comput Math Methods Med 2013;2013:890170. Cloutier G, Chen D, Durand LG. Performance of time-frequency representation techniques to measure blood flow turbulence with pulsed wave Doppler ultrasound. Ultrasound Med Biol 2001;4:535–550. Doi K, Giger ML, MacMahon H, Hoffmann KR, Nishikawa RM, Schmidt RA, Chua KG, Katsuragawa S, Nakamori N, Sanada S, et al. Computer-aided diagnosis: development of automated schemes for quantitative analysis of radiographic images. Semin Ultrasound CT MR 1992;13:140–152. Doi K. Computer-aided diagnosis in medical imaging: Historical review, current status and future potential. Comput Med Imaging Graph 2007;31:198–211. Doi K. Current status and future potential of computer-aided diagnosis in medical imaging. The Br J Radiol 2005;78:S3–S19. Doi K, MacMahon H, Katsuragawa S, Nishikawa RM, Jiang Y. Computer-aided diagnosis in radiology potential and pitfalls. Eur J Radiol 1999;2:97–109. Doi K. Computer-aided diagnosis in medical imaging achievements and challenges. In: World Congress on Medical Physics and Biomedical Engineering. Munich, Germany: 11th International Congress of the IUPESM; 2009. Giger ML. Computer-aided diagnosis of breast lesions in medical images. Comput Sci Eng 2000;2:39–45. Giger ML, Karssemeijer N. Computer-aided diagnosis in medical imaging. IEEE Trans Med Imaging 2001;20:1205–1208. Giger ML. Computerized analysis of images in the detection and diagnosis of breast cancer Seminars in Ultrasound. CT MRI 2004;25: 411–418. Lai YC, Huang YS, Wang DW, Tiu CM, Chou YH, Chang RF. Computer-aided diagnosis for 3-D power Doppler breast ultrasound. Ultrasound Med Biol 2013;39:555–567. Mayer D, Vomweg TW, Faber H, Weinheimer O, Mattiuzzi M, Buscema M, D€uber C. Fully automatic breast cancer diagnosis in contrast enhanced MRI. Int J Comp Assist Radiol Surg 2006;1: 325–343. Metz CE. Fundamental ROC analysis. In: Beutel J, Kundel HL, Van Metter RL, (eds). Handbook of medical imaging. Bellingham, Washington: SPIE Press, The International Society for Optical Engineering; 2000. p. 751–770. Nicolaides AN, Barnett JHM, Belcaro GV, Bernstein EF, Callow A, Eastcott HHG, Eikelboom B, Eklof B, Geroulakos G, Halliday A, Hobson R, Kalodiki E, Lusby R, Malikides A, Middleton LT, Moore WS, Norris J, Prescott L, Ramaswami G, Rutherford R, Sandmann W, Shanik G, Spartera C, Thomas D. Consensus statement on the management of patients with asymptomatic atherosclerotic carotid bifurcation lesions. Int Angiol 1995;14:5–17.
8
Ultrasound in Medicine and Biology
Parvati K, Prakasa Rao BS, Mariya Das M. Image segmentation using gray-scale morphology and marker-controlled Watershed transformation. Hindawi Publishing Corporation. Discrete Dynamics in Nature and Society 2008;2008:1–8. Perkins NJ, Schisterman EF, Vexler A. ROC curve inference from a sample with a limit of detection. Am J Epidemiol 2007;165:325–333. Petrick N, Chan HP, Wei D, Sahiner B, Helvie MA, Adler DD. Automated detection of breast masses on mammograms using adaptive contrast enhancement and texture classification. Med Phys 1996; 23:1685–1696. Rangaraj MR. Biomedical image analysis. CRC Press; 2005. Rossi AC, Brands PJ, Hoeks APG. Automatic recognition of the common carotid artery in longitudinal ultrasound B-mode scans. Med Image Analysis 2008;12:653–665. Seabra J, Pedro ML, Fernandes JF, Sanches J. Ultrasound Plaque Enhanced Activity Index for Predicting Neurologic Symptoms.
Volume -, Number -, 2014 Pattern Recognition and Image Analysis Lecture Notes in Computer Science 2011;2011(6669):184–191. Suetens P. Fundamentals of medical imaging. Cambridge University Press; 2002. Summers RM. Road maps for advancement of radiologic computeraided detection in the 21st century. Radiology 2003;229:11–13. Swets JA. Measuring the accuracy of diagnostic systems. Science 1988; 4857:1285–1293. Vincent L, Soille P. Watersheds in digital spaces: an efficient algorithm based on immersion simulations. IEEE Trans Pattern Anal Mach Intel 1991;6:583–598. Vyborny CJ, Giger ML, Nishikawa RM. Computer-aided detection and diagnosis of breast cancer. Radiol Clin North Am 2000;38: 725–740. Zweig MH, Campbell G. Receiver-Operating Characteristic (ROC) plots a fundamental evaluation tool in clinical medicine. Clin Chem 1993;19:561–577.
