- Email: [email protected]
Automated computerized scheme for detection of unruptured intracranial aneurysms in three-dimensional magnetic resonance angiography1
Automated Computerized Scheme for Detection of Unruptured Intracranial Aneurysms in ThreeDimensional Magnetic Resonance Angiography1 Hidetaka Arimura, PhD, Qiang Li, PhD, Yukunori Korogi, MD, Toshinori Hirai, MD, Hiroyuki Abe, MD, Yasuyuki Yamashita, MD, Shigehiko Katsuragawa, PhD, Ryuji Ikeda, RT, Kunio Doi, PhD
Rationale and Objectives. A computerized scheme for automated detection of unruptured intracranial aneurysms in magnetic resonance angiography was developed based on the use of a three-dimensional selective enhancement filter for dots (aneurysms). Materials and Methods. Twenty-nine cases with 36 unruptured aneurysms (diameter, 3 to 26 mm; mean, 6.6 mm) and 31 non-aneurysm cases were used in this study. The isotropic 3-dimensional magnetic resonance angiography images with 400 ⫻ 400 ⫻ 128 voxels (voxel size, 0.5 mm) were processed by use of the selective enhancement filter. The initial candidates were identified by use of a multiple gray-level thresholding technique on the dot-enhanced images and a regiongrowing technique with monitoring some image features. All candidates were classified into four types of candidates according to the size and local structures based on the effective diameter and skeleton image of each candidate (ie, large candidates and three types of small candidates including short-branch type, single-vessel type, and bifurcation type). In each group, a number of false-positives were removed by use of different rules on localized image features related to gray levels and morphology. Linear discriminant analysis was used for further removal of false-positives. Results. With this computer-aided diagnostic scheme, all of 36 aneurysms were correctly detected with 2.4 false-positives per patient based on a leave-one-out-by-patient test method. Conclusion. This computer-aided diagnostic system would be useful in assisting radiologists for the detection of intracranial aneurysms in magnetic resonance angiography. Key Words. Computer-aided diagnosis (CAD); unruptured intracranial aneurysm; magnetic resonance angiography (MRA); selective enhancement filter. ©
AUR, 2004
An intracranial aneurysm is a swelling along a blood vessel in the brain. Prospective autopsy and angiographic studies Acad Radiol 2004; 11:1093–1104 1 From the Department of Radiology, The University of Chicago, 5841 S Maryland Ave, Chicago, IL 60637 (H.A., Q.L., H.Abe, S.K., K.D.); the Department of Radiology, University Hospital of Occupational & Environmental Health, Kitakyushu, Japan (Y.K.); and the Department of Radiology, Kumamoto University Hospital, Kumamoto, Japan (T.H.,Y.Y., R.I.). Received April 1, 2004; revision requested June 15; accepted June 16. Supported by US Public Health Service grant nos. CA 62625 and CA 98119. K.D. and S.K. are shareholders in R2 Technology, Inc, Los Altos, CA. K.D. is a shareholder of Deus Technologies, Inc, Rockville, MD. Address correspondence to H.A., Kyushu University, Fukuoka 812-8582, Japan. e-mail: [email protected]
© AUR, 2004 doi:10.1016/j.acra.2004.07.011
indicated that between 3.6% and 6% of the general population have intracranial aneurysms (1), which could cause a subarachnoid hemorrhage (SAH) because of rupture of the aneurysm (2). Subarachnoid hemorrhage is a serious disorder with high mortality and morbidity (1,3– 6) (approximately 40% to 50% for mortality rate (3,7)). The rupture rate of asymptomatic aneurysms was estimated to be 1% to 2% per year (1,8). The accepted reference standard for identification of intracranial aneurysms is intra-arterial digital subtraction angiography (9 –11), which is invasive, time-consuming, and relatively expensive. During the past decade, there has been considerable interest in the roles of “less invasive” imaging modalities such as computed tomographic angiography
1093
ARIMURA ET AL
and magnetic resonance angiography (MRA) in the detection of intracranial aneurysms (1,9 –15). However, we believe that computed tomographic angiography should be considered a relatively invasive examination compared with MRA because patients are exposed to x-rays together with the injection of contrast medium. On the other hand, MRA can non-invasively detect unruptured intracranial aneurysms without the use of contrast media, at a performance level comparable to that by computed tomographic angiography (11). However, Korogi et al (15) investigated the diagnostic accuracy of MRA for detecting intracranial aneurysms on maximum intensity projection images with 78 aneurysms (including 60 [77%] ⬍5 mm in diameter) obtained from 61 patients. As a result, the sensitivities of aneurysms smaller than and larger than 5 mm in diameter were 56% and 86%, respectively. Moreover, White et al (11) reported results similar to those of Korogi et al (15), ie, 35% and 86% for the sensitivities of small and larger aneurysms (including 72 [67%] ⬍5 mm), respectively. Despite the recent advantages of MRA, it is still difficult and time-consuming for radiologists to find small aneurysms or it may not be easy to detect even medium-sized aneurysms on maximum intensity projection images because of overlapping with adjacent vessels and unusual locations. Therefore, a computer-aided diagnostic (CAD) scheme would be useful in assisting radiologists in the detection of intracranial aneurysms, especially small aneurysms, by use of MRA. Our purpose was to develop an automated computerized scheme for detection of intracranial aneurysms in MRA. Our CAD scheme was based on the use of a 3-dimensional (3D) selective enhancement filter (16) for dots, which would correspond to aneurysms in this study. The aneurysm candidates and false-positives were distinguished by analysis of localized image features and by use of rules based on size and local structures. The performance of our CAD scheme was evaluated by use of 31 non-aneurysm cases and 29 cases with 36 aneurysms of various sizes and at various locations.
MATERIALS AND METHODS Clinical Cases For evaluation of possible intracranial vascular disease, MRA studies of 60 patients were acquired on a 1.5 T magnetic resonance imaging scanner (Magnetom Vision; Siemens Medical Systems, Erlangen, Germany) by use of
1094
Academic Radiology, Vol 11, No 10, October 2004
Figure 1. Distribution of diameters of 36 unruptured aneurysms in MRA images.
a 3D time-of-flight technique in the Department of Radiology, Kumamoto University Hospital (Kumamoto, Japan). Each axial image was 512 ⫻ 512 pixels with a pixel size of 0.391 mm for 56 cases, and a pixel size of 0.410 mm for four cases. The 3D MRA images included 128 slices for 55 cases with a slice thickness of 0.5 mm, 96 slices for two cases with a slice thickness of 0.67 mm, and 64 slices for three cases with a slice thickness of 1.0 mm. All original 3D MRA images were converted to isotropic volume data that were used for training and testing in this study by use of linear interpolation and/or cropping, where each of the volume data was 400 ⫻ 400 ⫻ 128 voxels with a voxel size of 0.5 mm. The clinical cases used in this study consisted of 29 cases with 36 aneurysms (diameter measured by radiologists, 3 to 26 mm; mean, 6.6 mm) and 31 non-aneurysm cases. The 31 non-aneurysm cases included 26 normals and 5 abnormal cases with other vascular diseases (ie, old brain infarction, old brain hemorrhage, intracranial steno-occlusive disease, meningioma, pituitary microadenoma, and azygous anterior cerebral artery), whereas 12 of the 29 aneurysm cases also included other vascular diseases. Figure 1 shows the distributions of measured diameters for the 36 unruptured aneurysms. Approximately two thirds of all aneurysms were smaller than 6.0 mm, and one aneurysm was very large. Thirty-four aneurysms were saccular in shape, and two were fusiform. Figure 2 shows the distribution of aneurysms at various locations, which was considered similar to the distribution in a clinical environ-
Academic Radiology, Vol 11, No 10, October 2004
DETECTING UNRUPTURED ANEURYSMS IN 3D MRA
Figure 2. Distribution of 36 aneurysms at various locations, including internal carotid artery (ICA) and middle cerebral artery (MCA), anterior communicating artery (ACoA), basilar artery (BA), anterior cerebral artery (ACA), posterior cerebral artery (PCA), and vertebral artery (VA).
ment. Approximately two thirds of all aneurysms were found on the internal carotid artery and middle cerebral artery. Overall Scheme Figure 3 shows the overall CAD scheme developed in this study. First, the isotropic 3D MRA images were processed by use of three selective filters (16) for enhancement of aneurysms, vessels, and vessel walls. The initial candidates were identified by use of a multiple gray-level thresholding technique on the 3D dot-enhanced images within the search area, which was determined by dilation of major vessels (17), because most aneurysms appear on specific vessels, as shown in Figure 2. Candidate regions were segmented by use of a region-growing technique with monitoring some image features. In the next step, all candidates were classified into four types according to their size and local structures. In each group, a number of false positives were removed by use of rules based on localized image features related to gray levels and morphology. Finally, linear discriminant analysis was used for further removal of false positives. Selective Enhancement Filters Recently, Li et al (16) developed three selective enhancement filters based on the eigenvalues of a Hessian matrix with multiscales for dot, line, and plane, which can simultaneously enhance objects of specific shapes (such as dot-like lung nodules) and suppress objects of other shapes (such as line-like vessels). Because some aneurysms are round protrusions and others are bal-
Figure 3. Computerized scheme for automated detection of unruptured intracranial aneurysms in 3D MRA.
loon-like objects, which appear on intracranial vessels, many aneurysm shapes were hemispherical or spherical. Therefore, for enhancement of aneurysms and suppression of other objects such as vessels, the isotropic 3D MRA images were processed by use of the dotenhancement filter, and the dot-enhanced images were used for identification of initial aneurysm candidates and segmentation of candidate regions. Furthermore, the isotropic 3D images were processed by use of lineand plane-enhancement filters for “vessels” (ie, blood flow), and vessel walls (ie, surface of blood flow) for determination of localized image features (average voxel value and standard deviation [SD] of voxel value) of aneurysm candidates, because these image features for aneurysms would be different from those for non-aneurysms in the line- and plane-enhanced images. Figure 4 shows an original MRA image and three images selectively enhanced for dot, line, and plane objects, all of which were produced by maximum intensity projection image processing. In the dot-enhanced image (Fig 4b), an aneurysm was enhanced well and “vessels” disappeared, although some “nonaneurysms” were also enhanced, which included bending regions and vessel bifurcations. On the other hand, the aneurysm disappeared in the line-enhanced image
1095
ARIMURA ET AL
Figure 4. Illustration of (a) an original MRA image and three images selectively enhanced for (b) dot, (c) line, and (d) plane objects, all of which were produced by maximum intensity projection image processing. Arrows indicate a large (7.5 mm) aneurysm.
(Fig 4c), but most of the vessels remained, and the walls of the aneurysm and vessels were enhanced in the plane-enhanced image (Fig 4d). Determination of Initial Aneurysm Candidates Identification of Initial Candidates Based on Multiple Gray-Level Thresholding For identification of initial aneurysm candidates, a multiple gray-level thresholding technique was applied to the dot-enhanced image that was smoothed by averaging with a square kernel (3 ⫻ 3 ⫻ 3) to reduce noise. Each threshold level was determined according to a certain spe-
1096
Academic Radiology, Vol 11, No 10, October 2004
cific upper percentage of the area under the pixel-value histogram in the dot-enhanced image within the search area. The pixel values of aneurysms in the dot-enhanced image were usually located at the high end of its histogram (the portion of the histogram closest to the highest pixel value), which ranged from approximately 0.008% to 0.8% for the cases used in this study. The regions in the dot-enhanced image above a certain threshold value were called “islands” (3D objects in 3D space), which were initial candidates. Note that the area under the histogram for each threshold level is equivalent to the total volume of all islands emerged. For initially picking up as many aneurysms as possible, the island volume and corresponding threshold level should be increased by a small enough volume for detecting a small aneurysm. Therefore, the incremental percentage of the area under the histogram for each threshold level was determined empirically by a fraction of a small volume relative to the total volume of the search area, which is equivalent to the total area of the histogram (eg, a small volume of approximately 20 mm3 and 85 mm3; the corresponding cube size of approximately 2.7 mm and 4.4 mm) for higher and lower threshold levels, respectively. At the first percentage threshold level where each island emerged (referred to as “starting percentage threshold level”), the effective diameter was determined for selection of initial aneurysm candidates. The effective diameter of a candidate was defined by the diameter of a sphere with the same volume as that of the candidate. If the effective diameter of an island was greater than 2 mm at the starting percentage threshold level, the island was considered an initial aneurysm candidate. Thus, an initial candidate selected at a starting percentage threshold level would not be examined again at the subsequent percentage threshold levels. Determination of Candidate Regions by Use of RegionGrowing Technique on Dot-Enhanced Image For each of the initial candidates, its region was determined by application of a region-growing technique to the dot-enhanced image to obtain the image features of the candidates for subsequent rule-based schemes. The candidate regions were determined within a volume of interest (40 mm ⫻ 40 mm ⫻ 40 mm), where the center of the volume of interest was located at the voxel with a maximum value for each initial candidate in the dot-enhanced image. Our segmentation method of the candidate region was based on finding a large change (referred to as a
Academic Radiology, Vol 11, No 10, October 2004
“transition point”) in some image features, which implied that the candidate region merged with its adjacent background structures or other candidates, as the candidate region grew. The region growing began at the location where the voxel value was the maximum in the initial candidate region, and was repeated at various gray levels, which were decreased from each previous gray level with a decrement of 5% of the maximum voxel value. The percentage of a gray-level decrease from the maximum value, which was referred to as “percentage gray level,” changed from 5% to 90%. At each percentage gray level, in this study, two image features for the candidate grown region (ie, the effective diameter and average contrast) were determined for finding the transition point in these features. The average contrast was defined by the difference in the average voxel values between the inside and outside regions in the original image divided by the average voxel value of the inside region. The inside and outside regions for each candidate were defined by the candidate region at a current percentage gray level and the increased region at the subsequent percentage gray level, respectively. However, for candidates without transition points in the effective diameter and the average contrast, the candidate regions were determined at the 90% gray level. Localized Feature Analysis and False Positive Removal Based on Rules According to Size and Local Structures All initial candidates were grouped into small and large candidates based on their effective diameters, some of the small and large false positives were then removed by use of different rules in each group (the first rulebased scheme). The remaining small candidates were further classified into three groups according to the local structures based on the skeleton image, which included a short-branch type, a single-vessel type, and a bifurcation type (including trifurcation), and some of the three types of false positives were removed by use of another set of different rules in each group (the second rule-based scheme). Classification of Initial Candidates Based on Effective Diameter It is important to note that the characteristics of large aneurysms in MRA images were quite different from those of small aneurysms. For example, the voxel values in the core region (near the center) for large aneurysms were commonly lower than those in the rind region adjacent to the
DETECTING UNRUPTURED ANEURYSMS IN 3D MRA
aneurysm wall because of the slow speed or turbulence of blood flow inside the aneurysm. By observing the original axial images of many aneurysms, we found that aneurysms with diameters larger than approximately 6.5 mm had this unique characteristic. In this study, initial candidates were grouped into small and large candidates by use of an effective diameter of 6.5 mm. First Rule-Based Schemes for Removal of Small and Large False Positives For the initial removal of small and large false positives, we used the image features of the gray level, size, and shape. Generally, the sizes of some false positives were smaller or larger than those of aneurysms, and some were less circular or more irregular compared with aneurysms. The average voxel values, SDs, and contrasts of some false positives were smaller or larger than those of aneurysms because of slow blood flow in some small vessels or fast speed in the bifurcation or bending regions on some vessels, respectively, and also for some other reasons caused by the turbulence of the blood flow inside some aneurysms and the nonuniform blood flow speed inside some vessels with nonuniform diameters. Therefore, we determined the gray-level features in the dot-enhanced and original images, ie, the average voxel value, the relative SD of the voxel value, the relative contrast, the average contrast (defined previously), the relative difference in the SD of voxel values between the candidate and outside regions, and the morphologic features, ie, the effective diameter, the sphericity, the relative SD of the distance between the centroid and the surface, and the maximum and minimum distance between the centroid and the surface. The “relative value” means the value relative to the average value. The degree of sphericity was defined by the fraction of the overlap volume between the candidate region and the sphere (with the same volume as the candidate volume). The relative SDs of voxel values were obtained in both the candidate and outside regions because local structures inside and outside aneurysms could differ from those for false positives. The relative contrast used in this study was defined by the difference between the maximum and minimum voxel values within the candidate region divided by the average voxel value. The average voxel values and the relative SD of the voxel values were determined in the line- and plane-enhanced images as well, because, as shown in Figure 4, aneurysms almost disappeared, but vessels were enhanced well in the line-enhanced images, and the appearances inside the aneurysm surfaces were different
1097
ARIMURA ET AL
from those of the non-aneurysm vessels in the plane-enhanced images. The relative SD of the distance between the centroid and the surface was related to the degree of irregularity for the surface of the candidate region. In addition, for removal of large false-positives, we determined specific features relevant to the characteristics of large aneurysms, ie, the average voxel values in the core and rind regions on the original image, the relative SDs of the voxel values in both regions, the relative differences in the average voxel values, and the SDs between the core and rind regions. All of the rules used in the two rule-based schemes were based on removal of false positives by use of simple thresholding for both the upper and lower limits of the features determined from all aneurysms included in each group; upper and lower limits were obtained, respectively, as 5% higher and 5% lower than the maximum and minimum values of each feature. If one of the features for a candidate was larger than the upper limit or smaller than the lower limit, the candidate was removed as a false positive. Classification of Small Candidates According to Local Structures Based on Skeleton Images It should be noted that the image features for small aneurysms and non-aneurysms were different in each group because of differences in the types of local structures. In addition, candidates with the short branch in a skeleton image (referred to as “short-branch type”) should be examined carefully because this short branch could be an aneurysm in an original image; details are explained in the Appendix. Therefore, we classified small candidates into three groups (short-branch type, single-vessel type, and bifurcation type), and established a number of effective rules to remove many non-aneurysms in each group. The grouping was made according to the local structures based on the skeleton image, which was obtained from the distance-transformed image, because the topologic properties (based on connectivity) would be preserved in the skeleton image (18,19). For example, the structure of a vessel would be simplified by maintaining voxels with only 1 voxel width. The local structure of the candidate was determined by counting the number of the skeleton objects in a rind region (a rind thickness of 1.0 mm) of a sphere of diameter 1.4 times larger than the candidate region. For example, if the number of skeleton objects was one or two, the local structure would be a vessel end or a single vessel, respectively. The details on the method
1098
Academic Radiology, Vol 11, No 10, October 2004
for classification of small candidates are described in the Appendix. Second Rule-Based Schemes for Removal of Small FalsePositives In the second rule-based scheme for further removal of many small false-positives (more than 80% false positives) for each group of the candidates, we established respective rules based on the localized image features. In this rule-based scheme, additional localized features were determined in both the distance-transformed image and the candidate region segmented in the original image. Note that the distance value in the distance-transformed image was related to the thickness of the vessel or the diameter of the vessel cross-section. In general, because small aneurysms may be considered as protrusions on the vessels, the diameter of the vessel cross-section with the aneurysm would be larger than the vessel without the aneurysm. Therefore, the maximum distance values in the candidate and outside regions, and the average distance values in the outside regions were obtained from the distance-transformed image, and also the relative difference in the maximum distance values (and the average distance values) between the candidate and the outside regions were determined. For the segmentation of the candidate region in the original image, we applied a region-growing technique as previously described. However, we found that not only the regions of the aneurysms (and false-positives) but also connected vessels were segmented at this stage. Therefore, the candidate region in the original image was segmented by region-growing only within the dilated volume of the candidate region obtained in the dot-enhanced image. However, for candidates of the short-branch type, the dilated volume was derived by dilation of the skeleton image of the short branch, because the candidates of the short-branch type were found based on the skeleton image, not the dot-enhanced image. The morphologic features for the candidate region in the original image included the effective diameter, the sphericity, the relative SD of the distance between the centroid and the surface, and the maximum and minimum distances between the centroid and the surface, which were the same as those used for the candidate region obtained in the dot-enhanced image. Additional features were determined for the short-branch type candidates (ie, the protrusion length and the average distance value in the candidate region obtained from the distance-transformed image). All of the rules used in the second rule-based scheme
Academic Radiology, Vol 11, No 10, October 2004
DETECTING UNRUPTURED ANEURYSMS IN 3D MRA
were based on removal of false positives by use of the same simple thresholding (upper and lower limits) as those used in the first rule-based scheme described previously. Linear Discriminant Analysis In our scheme, 43 image features can be used for the linear discriminant analysis for further removal of false positives. However, because the short-branch type candidates were detected only in the second step for the removal of small false positives, the number of features for the short-branch type candidates were limited (ie, the average voxel value, the relative SD of the voxel value, the relative contrast in the original image, and the effective diameter, the sphericity, the relative SD of the distance between the centroid and the surface, and the maximum and minimum distances between the centroid and the surface). Because some of the features were not useful for classification of the remaining candidates as aneurysms or false positives by use of a linear discriminant function, we selected the most effective combination of image features based on Wilks’ lambda and on the Az value (ie, the area under the receiver operating characteristic [ROC] curve) (20,21). Consequently, the final combination consisted of four features (ie, the average voxel value, the relative SD of the voxel value, the relative SD of the distance between the centroid and the surface, and the difference between the maximum and minimum distance [between the centroid and the surface]). Evaluation of the Performance Our CAD scheme was evaluated with both a consistency test and a leave-one-out-by-patient test method. With the consistency test, cases used for training were also used for testing. With the leave-one-out-by-patient test method, all candidates except one and candidates obtained from the same patient were used for training, and the one candidate left out was used for testing with two rule-based schemes and the linear discriminant function. This procedure was repeated for all candidates, so that each candidate was used once as a test candidate. By changing a threshold value for discriminant scores of the candidates produced with the linear discriminant function, which was determined by linear discriminant analysis for distinction between aneurysms and false positives, we determined the free-response receiver operating characteristic (FROC) curve of the CAD scheme in the leave-oneout-by-patient test method.
Figure 5. Results obtained with the segmentation of (a) small and (b) large aneurysms by use of the region-growing technique on the dot-enhanced or original images.
RESULTS AND DISCUSSION The computerized scheme for detection of the intracranial aneurysms in MRA was applied to 31 non-aneurysm cases and 29 abnormal cases with 36 aneurysms. All of the 36 aneurysms with 22.3 false positives per patient were detected at the initial identification step. For evaluation of our scheme in the computerized detection of aneurysms, we used a criterion such that an aneurysm was considered correctly detected if the location of the maximum voxel value in the candidate region was within the diameter of the aneurysm measured by radiologists for aneurysms smaller than 7.0 mm, and within the diameter of 7.0 mm for aneurysms larger than 7.0 mm. Figure 5 shows the results obtained with the segmentation of small and large aneurysms by use of the region-
1099
ARIMURA ET AL
Academic Radiology, Vol 11, No 10, October 2004
Figure 6. Relationship between the effective diameter in the dot-enhanced image and the relative SD of the voxel value in the original image for small and large aneurysms (circles and squares) as well as small and large false positives (dots and triangles) in the dot-enhanced image. Rules are indicated by dashed lines.
Figure 7. Relationship between the maximum distance value in the candidate region and average distance value in the outside region, for single-vessel-type candidates (aneurysms: circles; false positives: dots). Rules are indicated by dashed lines.
growing technique on the dot-enhanced or original images, where most of the small and large aneurysms were segmented well. However, we found that segmentation of large aneurysms tended to be less accurate; not only the aneurysm, but also adjacent vessel regions (or background) were included in the segmented candidate region, (eg, large aneurysms in the middle and bottom of Figure 5b). These inaccuracies of segmentation occurred because the contrasts of the surface area of such large aneurysms were very low, and the average voxel values in the core region for the aneurysms were lower than those in the rind region. Nevertheless, it was not difficult to distinguish between large aneurysms and large false positives because of the unique image features of large aneurysms. In the first rule-based scheme, all initial candidates were grouped into small and large candidates, and many false positives were removed by use of rules based on the localized image features. Figure 6 shows the relationship between the effective diameter in the dot-enhanced image and the relative SD of the voxel value in the original image. The effective diameters of some false positives, such as small short vessels and large long vessels, were smaller and larger, respectively, than those of aneurysms. In addition, because the distribution of voxel values inside some vessels (eg, vessels of elderly patients with
stenosis and occlusion) were more nonuniform than those of aneurysms, the relative SDs of the voxel values for these false positives were greater than those of aneurysms. Therefore, many false positives were removed by use of such rules as shown by dashed lines in Figure 6. By use of the first rule-based scheme, all aneurysms were retained, and the average number of false positives per patient was reduced from 22.3 to 5.8. At this stage, the majority of the remaining false positives were due to bending of single vessels and bifurcation vessels, most of which were further removed by use of the second rulebased scheme, as described below. For single-vessel-type candidates, Figure 7 shows the relationship between the average distance values in the outside region and the maximum distance values in the candidate region, where the distance values were obtained from the distance-transformed images. Note that many bending regions of large vessels were included as falsepositives in single-vessel-type candidates, and the average distance values in the outside region for such false-positives were larger than those of aneurysms, as shown by many false positives on the right side in Figure 7. In addition, the difference in the distance values between the candidate region and the outside region tended to be larger for aneurysms than those for most of the false posi-
1100
Academic Radiology, Vol 11, No 10, October 2004
Figure 8. FROC curve for overall performance of our computerized scheme in automated detection of intracranial aneurysms in MRA images.
tives. Finally, many false positives in each group were removed based on the differences in the image features between aneurysms and false positives by use of rules as shown by dashed lines in Figure 7. As a result, all of the 36 aneurysms were detected correctly with 0.55 falsepositives per patient in a consistency test. Furthermore, as a result of the leave-one-out-by-patient test method, our scheme achieved a sensitivity of 100% with 2.4 falsepositives per patient. Figure 8 shows the FROC curve for the overall performance of our scheme by use of linear discriminant analysis with the leave-one-out-by-patient test method. According to this result, our CAD scheme had a sensitivity of 89% with 1.3 false-positives per patient. At this operating point, four aneurysms were not detected by our CAD scheme, and thus may be considered difficult cases for a computerized detection. These cases were two large aneurysms and two short-branch type small aneurysms. Segmentation of the large aneurysms could be inaccurate as shown in Figure 5, because the contrasts of the surface area of large aneurysms were very low and the average voxel values in the core region for large aneurysms were lower than those in the rind region. The segmented regions of two large aneurysms included adjacent vessels and some cavities. Short-branch type small aneurysms adjacent to large regions, especially the bending region of a large parent vessel, were hardly enhanced in the dot-
DETECTING UNRUPTURED ANEURYSMS IN 3D MRA
enhanced image, because such large regions tended to be strongly enhanced by the dot-enhancement filter. Although some of small aneurysms were detected by finding the short branches, the segmentations for some of shortbranch type aneurysms tended to be relatively inaccurate. Therefore, we need to improve the segmentation of large aneurysms and short-branch type aneurysms in the future. During the last two decades, we developed a number of CAD schemes for detection and classification of various abnormalities such as microcalcifications and masses in mammograms (22,23), pulmonary nodules and interstitial infiltrates in chest radiographs (24,25), and nodules and diffuse lung diseases in computed tomography (26,27). We have shown the usefulness of these schemes by carrying out a number of observer performance studies (28,29). Therefore, for investigating the usefulness of our CAD scheme for detection of intracranial aneurysms in MRA, it will be also necessary to carry out an observer study by use of ROC analysis to compare a radiologists’ performance without and with the aid of CAD output in terms of the detection accuracy of aneurysms in MRA. Our database included many different kinds of cases with aneurysms that were selected by considering the size, shape, and location of aneurysms, the gender and age of the patients, and other vascular diseases such as stenosis or occlusion. However, because all cases used in this study were obtained from a magnetic resonance imaging scanner in one hospital, our results would depend on the magnetic resonance imaging scanner and the acquisition sequence used in the study. Furthermore, the number of the false positives or specificity would depend on the patients with other vascular diseases, because the appearance of the blood flow within such vessels in the MRA images might look like some aneurysms for the computer. Therefore, our CAD scheme needs to be applied to an independent database including many cases acquired from different magnetic resonance imaging scanners on many different patients. In addition, because all rules used in this study were determined based on image features of a relatively small number of aneurysms (ie, from 6 to 11 in each group), the number of abnormal cases needs to be increased for determination of more general and robust rules for unknown cases. It has been reported that screening of intracranial aneurysms would be especially useful in high-risk groups (eg, patients with adult polycystic kidney disease or a strong family history of aneurysm SAH) (30 –33). Magnetic resonance angiography may be used as screening examinations in detecting intracranial aneurysms, where our CAD
1101
ARIMURA ET AL
scheme could aid radiologists in detecting aneurysms; this may be similar to the usefulness of the CAD scheme in the detection of breast cancers in screening mammograms (34). In conclusion, we have developed a CAD scheme based on the use of selective enhancement filters for the automated detection of unruptured intracranial aneurysms in MRA. With our CAD scheme, all of the 36 aneurysms in 60 cases were detected correctly with 0.55 false-positives per patient in an evaluation by a consistency test, and 2.4 false-positives by the leave-one-out-by-patient test method. Our CAD system would be useful in assisting radiologists in the detection of unruptured intracranial aneurysms in MRA.
Academic Radiology, Vol 11, No 10, October 2004
14. 15.
16.
17.
18.
19.
ACKNOWLEDGMENT
The authors thank Junji Shiraishi, PhD, Kenji Suzuki, PhD, Feng Li, MD, Chisako Muramatsu, BS, and Roger Engelmann, MS, for their useful discussions, and Elisabeth Lanzl for improving the manuscript.
20. 21.
22.
REFERENCES 1. Wardlaw JM, White PM. The detection and management of unruptured intracranial aneurysms. Brain 2000; 123:205–221. 2. Weir B. Unruptured intracranial aneurysms: A review. J Neurosurg 2002; 96:3– 42. 3. Fogelholm R, Hernesniemi J, Vapalahti M. Impact of early surgery on outcome after aneurysm subarachnoid hemorrhage: a populationbased study. Stroke 1993; 24:1649 –1654. 4. Hop JW, Rinkel GJ, Algra A, van Gijn J. Case-fatality rates and functional outcome after subarachnoid hemorrhage: A systematic review [review]. Stroke 1997; 28:660 – 664. 5. Hijdra A, Braakman R, van Gijn J, Vermeulen M, van Crevel H. Aneurysmal subarachnoid hemorrhage: Complications and outcome in a hospital population. Stroke 1987; 18:1061–1067. 6. Sengupta RP, McAllister VL. Subarachnoid hemorrhage. Berlin: SpringerVerlag, 1986. 7. Broderick JP, Brott TG, Duldner JE, Tomsick T, Leach A. Initial and recurrent bleeding are the major causes of death following subarachnoid hemorrhage. Stroke 1994; 25:1342–1347. 8. Juvela S, Porras M, Poussa K. Natural history of unruptured intracranial aneurysms: Probability of and risk factors for aneurysm rupture. J Neurosurg 2000; 93:379 –387. 9. Ogawa T, Okudera T, Noguchi K, Sasaki N, Inugami A, Uemura K, Yasui N. Cerebral aneurysms : Evaluation with three-dimensional CT angiography. AJNR Am J Neuroradiol 1996; 17:447– 454. 10. Korogi Y, Takahashi M, Katada K, Ogura Y, Hasuo K, Ochi M, Utsunomiya H, Abe T, Imakita S. Intracranial aneurysms: Detection with three-dimensional CT angiography with volume rendering – Comparison with conventional angiographic and surgical findings. Radiology 1999; 211:497–506. 11. White PM, Teasdale EM, Wardlaw JM, Easton V. Intracranial aneurysms: CT angiography and MR angiography for detection – Prospective blinded comparison in a large patient cohort. Radiology 2001; 219:739 –749. 12. Preda L, Gaetani P, Rodriguez y Baena R, et al. Spiral CT angiography and surgical corrections in the evaluation of intracranial aneurysms. Eur Radiol 1998; 8:739 –745. 13. Huston J III, Nichols DA, Luetmer PH, Goodwin JT, Meyer FB, Wiebers DO, Weaver AL. Blinded prospective evaluation of sensitivity of MR
1102
23. 24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
angiography to known intracranial aneurysms: Importance of aneurysm size. AJNR Am J Neuroradiol 1994; 15:1607–1614. Aprile I. Evaluation of cerebral aneurysms with MR-angiography. Riv Neuroradiol 1996; 9:541–550. Korogi Y, Takahashi M, Mabuchi N, et al. Intracranial aneurysms: Diagnostic accuracy of three-dimensional, Fourier transform, time-o-flight MR angiography. Radiology 1994; 193:181–186. Li Q, Sone S, Doi K. Selective enhancement filters for nodules, vessels, and airway wall in two- and three-dimensional CT scans. Med Phys 2003; 30:2040 –2051. Arimura H, Li Q, Korogi Y, et al. Development of CAD scheme for automated detection of intracranial aneurysms in magnetic resonance angiography. 18th International Congress of Computer Assisted Radiology and Surgery. Chicago, IL; International Congress Series 2004; 1268:1015–1020. Saito T, Toriwaki J. New algorithms for Euclidean distance transformations of an n-dimensional digitized picture with applications. Pattern Recognition 1994; 27:1551–1565. Saito T, Toriwaki J. A sequential thinning algorithm for three dimensional digital pictures using the Euclidean distance transformation. Proceedings of the 9th Scandinavian Conference on Image Analysis SCIA 95. Uppsala, Sweden 1995; 507–516. Johnson RA, Wichern DW. Applied Multivariate Statistical Analysis. Cliffs, NJ; Prentice-Hall, 1992, 5.3:184 –188. Aoyama M, Li Q, Katsuragawa S, et al. Automated computerized scheme for distinction between benign and malignant solitary pulmonary nodules on chest images. Med Phys 2002; 29:701–708. Chan HP, Doi K, Galhotra S, Vyborny CJ, MacMahon H, Jokich PM. Image feature analysis and computer-aided diagnosis in digital radiography. 1. Automated detection of microcalcifications in mammography Med Phys 1987; 14:538 –548. Kupinski MA, Giger ML. Automated seeded lesion segmentation on digital mammograms. IEEE Trans Med Imaging 1998; 17:510 –517. Giger ML, Doi K, MacMahon H. Image feature analysis and computeraided diagnosis in digital radiography. 3. Automated detection of nodules in peripheral lung fields Med Phys 1988; 15:158 –166. Katsuragawa S, Doi K, MacMahon H. Image feature analysis and computer-aided diagnosis in digital radiography. Detection and characterization of interstitial lung disease in digital chest radiographs Med Phys 1988; 15:311–319. Suzuki K, Armato SG III, Li F, Sone S, Doi K. Massive training artificial neural network (MTANN) for reduction of false positives in computerized detection of lung nodules in low-dose CT. Med Phys 2003; 30: 1602–1617. Uchiyama Y, Katsuragawa S, Abe H, et al. Quantitative computerized analysis of diffuse lung disease in high-resolution computed tomography. Med Phys 2003; 30:2440 –2464. Chan HP, Doi K, Vyborny CJ, et al. Improvement in radiologists’ detection of clustered microcalcifications on mammograms – The potential of computer-aided diagnosis. Invest Radiol 1990; 25:1102–1110. Shiraishi J, Abe H, Engelmann R, Aoyama M, MacMahon H, Doi K. Computer-aided diagnosis for distinction between benign and malignant solitary pulmonary nodules in chest radiographs: ROC analysis of radiologists’ performance. Radiology 2003; 227:469 – 474. Levey AS. Screening for occult intracranial aneurysms in polycystic kidney disease: Interim guidelines [editorial]. J Am Soc Nephrol 1990; 1:9 –12. Wiebers DO, Torres VE. Screening for unruptured intracranial aneurysms in autosomal dominant polycystic kidney disease [editorial]. N Engl J Med 1992; 327:953–955. Ronkainen A, Puranen MI, Hernesniemi JA, et al. Intracranial aneurysms: MR angiographic screening in 400 asymptomatic individuals with increased familial risk. Radiology 1995; 195:35– 40. Kojima M, Nagasawa S, Lee YE, Takeichi Y, Tsuda E, Mabuchi N. Asymptomatic familial cerebral aneurysms. Neurosurgery 1998; 43:776 –781. Freer TW, Ulissey MJ. Screening mammography with computer-aided detection: Prospective study of 12,860 patients in a community breast center. Radiology 2001; 220:781–786.
Academic Radiology, Vol 11, No 10, October 2004
Figure A1. Schematic diagram for classification of small candidates into three groups including short-branch type, singlevessel type, and bifurcation type based on the skeleton image.
APPENDIX The schematic diagram for classification of small candidates into three groups based on the skeleton image is shown in Figure A1. First, the candidate region was segmented within the volume of interest (30 mm ⫻ 30 mm ⫻ 30 mm) in the original image, where the centroid of the candidate region was located at the center of the volume of interest by use of the region-growing technique, which is similar to that used for the dotenhanced image. The distance-transformed image was derived by calculation of a Euclidean minimum distance from each voxel in the segmented region to the nearest background in the binary image (18). Next, the skeleton image was obtained by use of a thinning algorithm (19) based on the distance-transformed image, where deletable voxels with smaller distance values were removed first so that the topologic properties of the segmented vessels could be preserved. For classification of the local structures, the skeleton images of candidates were analyzed as shown in Figure A1. Note that “lump” candidates and short-branch type candidates were identified in this classification scheme, as described in detail in the next two paragraphs. If the degree of lump for the candidate (defined in the next paragraph) was smaller than 0.1, the number of skele-
DETECTING UNRUPTURED ANEURYSMS IN 3D MRA
ton objects was counted in the rind region. If not, candidates were classified as the bifurcation type, because most of the aneurysms with a degree of lump greater than 0.1 were located at the bifurcation. If the number of skeleton objects in the rind region was 1, the local structure would be considered as a single vessel. If the number of the skeleton objects was ⱖ2, the nearest short branch was searched in the large sphere including the original candidate. If a short branch was found, then the candidate was determined as the short-branch type. If the short branch was not found, and if the number of the skeleton objects was equal to 2, the candidate would be a single-vessel type; if the number of skeleton objects was ⱖ3, the candidate would be a bifurcation type. The skeleton objects for some candidates would have a “lump” composed of many short “skeletons,” which look like short hairs and would not be true skeletons; therefore, the number of the skeleton objects in the rind region could be incorrect. Thus, before counting the skeleton objects in the rind region, we determined the degree of lump for each candidate for identifying the lump candidates, where the degree of lump was defined by a fraction of the total volume of skeleton objects in a 2.0-mm sphere placed at the centroid of the candidate region. A small protrusion, a small branching vessel (nonaneurysm), and a small aneurysm on a single vessel or bifurcation (parent vessel) in the original image may be considered as a short branch attached to a parent skeleton in the skeleton image, as shown in Figure A2. Such candidates with a short branch were referred to as short-branch type and should be examined carefully because this short branch could be an aneurysm. Note that some small aneurysms adjacent to large regions, such as the bending region of a large parent vessel, were hardly enhanced in the dot-enhanced image because such large regions tended to be strongly enhanced by the dot-enhancement filter. Such small aneurysms could be detected by finding the short branch adjacent to the original candidate region in the skeleton image. Thus, the short branch was searched as illustrated in Figure A2: (1) the nearest short branch with an end point was searched within a sphere of diameter four times larger than the original candidate region, and the large sphere was placed at the centroid of the original candidate region; (2) if the short-branch length was larger than the radius of a parent vessel, and if the difference between the length and the radius was in the
1103
ARIMURA ET AL
Academic Radiology, Vol 11, No 10, October 2004
range from 1.5 mm to 6.5 mm, then the short branch could be a small protrusion on the parent vessel or a small aneurysm; and thus, (3) the short branch was considered as a new candidate of the short-branch type at this stage. Otherwise, the candidate was classified as either the bifurcation type or the single-vessel type. The difference between the short-branch length and the radius of a parent vessel was defined as the protrusion length, which was obtained as an image feature for the short-branch type candidates used in the second rulebased scheme. Because both the short branch and the vessel end have an end point, it was difficult to distinguish between them. Therefore, we decided not to examine whether the vessel-end candidates were of the short-branch type, and the vessel-end candidates were classified as the single-vessel type in this study, because a vessel end would consist of a single vessel. Figure A2. Illustration of a method for searching for a short branch adjacent to the original candidate region in the skeleton image.
The Association of University Radiologists wishes to thank the following companies for their generous support of the AUR Junior Membership Program:
1104
4th Year Radiology Residents
3rd Year Radiology Residents
Siemens Medical Systems, Inc.
Berlex Laboratories
Rationale and Objectives. A computerized scheme for automated detection of unruptured intracranial aneurysms in magnetic resonance angiography was developed based on the use of a three-dimensional selective enhancement filter for dots (aneurysms). Materials and Methods. Twenty-nine cases with 36 unruptured aneurysms (diameter, 3 to 26 mm; mean, 6.6 mm) and 31 non-aneurysm cases were used in this study. The isotropic 3-dimensional magnetic resonance angiography images with 400 ⫻ 400 ⫻ 128 voxels (voxel size, 0.5 mm) were processed by use of the selective enhancement filter. The initial candidates were identified by use of a multiple gray-level thresholding technique on the dot-enhanced images and a regiongrowing technique with monitoring some image features. All candidates were classified into four types of candidates according to the size and local structures based on the effective diameter and skeleton image of each candidate (ie, large candidates and three types of small candidates including short-branch type, single-vessel type, and bifurcation type). In each group, a number of false-positives were removed by use of different rules on localized image features related to gray levels and morphology. Linear discriminant analysis was used for further removal of false-positives. Results. With this computer-aided diagnostic scheme, all of 36 aneurysms were correctly detected with 2.4 false-positives per patient based on a leave-one-out-by-patient test method. Conclusion. This computer-aided diagnostic system would be useful in assisting radiologists for the detection of intracranial aneurysms in magnetic resonance angiography. Key Words. Computer-aided diagnosis (CAD); unruptured intracranial aneurysm; magnetic resonance angiography (MRA); selective enhancement filter. ©
AUR, 2004
An intracranial aneurysm is a swelling along a blood vessel in the brain. Prospective autopsy and angiographic studies Acad Radiol 2004; 11:1093–1104 1 From the Department of Radiology, The University of Chicago, 5841 S Maryland Ave, Chicago, IL 60637 (H.A., Q.L., H.Abe, S.K., K.D.); the Department of Radiology, University Hospital of Occupational & Environmental Health, Kitakyushu, Japan (Y.K.); and the Department of Radiology, Kumamoto University Hospital, Kumamoto, Japan (T.H.,Y.Y., R.I.). Received April 1, 2004; revision requested June 15; accepted June 16. Supported by US Public Health Service grant nos. CA 62625 and CA 98119. K.D. and S.K. are shareholders in R2 Technology, Inc, Los Altos, CA. K.D. is a shareholder of Deus Technologies, Inc, Rockville, MD. Address correspondence to H.A., Kyushu University, Fukuoka 812-8582, Japan. e-mail: [email protected]
© AUR, 2004 doi:10.1016/j.acra.2004.07.011
indicated that between 3.6% and 6% of the general population have intracranial aneurysms (1), which could cause a subarachnoid hemorrhage (SAH) because of rupture of the aneurysm (2). Subarachnoid hemorrhage is a serious disorder with high mortality and morbidity (1,3– 6) (approximately 40% to 50% for mortality rate (3,7)). The rupture rate of asymptomatic aneurysms was estimated to be 1% to 2% per year (1,8). The accepted reference standard for identification of intracranial aneurysms is intra-arterial digital subtraction angiography (9 –11), which is invasive, time-consuming, and relatively expensive. During the past decade, there has been considerable interest in the roles of “less invasive” imaging modalities such as computed tomographic angiography
1093
ARIMURA ET AL
and magnetic resonance angiography (MRA) in the detection of intracranial aneurysms (1,9 –15). However, we believe that computed tomographic angiography should be considered a relatively invasive examination compared with MRA because patients are exposed to x-rays together with the injection of contrast medium. On the other hand, MRA can non-invasively detect unruptured intracranial aneurysms without the use of contrast media, at a performance level comparable to that by computed tomographic angiography (11). However, Korogi et al (15) investigated the diagnostic accuracy of MRA for detecting intracranial aneurysms on maximum intensity projection images with 78 aneurysms (including 60 [77%] ⬍5 mm in diameter) obtained from 61 patients. As a result, the sensitivities of aneurysms smaller than and larger than 5 mm in diameter were 56% and 86%, respectively. Moreover, White et al (11) reported results similar to those of Korogi et al (15), ie, 35% and 86% for the sensitivities of small and larger aneurysms (including 72 [67%] ⬍5 mm), respectively. Despite the recent advantages of MRA, it is still difficult and time-consuming for radiologists to find small aneurysms or it may not be easy to detect even medium-sized aneurysms on maximum intensity projection images because of overlapping with adjacent vessels and unusual locations. Therefore, a computer-aided diagnostic (CAD) scheme would be useful in assisting radiologists in the detection of intracranial aneurysms, especially small aneurysms, by use of MRA. Our purpose was to develop an automated computerized scheme for detection of intracranial aneurysms in MRA. Our CAD scheme was based on the use of a 3-dimensional (3D) selective enhancement filter (16) for dots, which would correspond to aneurysms in this study. The aneurysm candidates and false-positives were distinguished by analysis of localized image features and by use of rules based on size and local structures. The performance of our CAD scheme was evaluated by use of 31 non-aneurysm cases and 29 cases with 36 aneurysms of various sizes and at various locations.
MATERIALS AND METHODS Clinical Cases For evaluation of possible intracranial vascular disease, MRA studies of 60 patients were acquired on a 1.5 T magnetic resonance imaging scanner (Magnetom Vision; Siemens Medical Systems, Erlangen, Germany) by use of
1094
Academic Radiology, Vol 11, No 10, October 2004
Figure 1. Distribution of diameters of 36 unruptured aneurysms in MRA images.
a 3D time-of-flight technique in the Department of Radiology, Kumamoto University Hospital (Kumamoto, Japan). Each axial image was 512 ⫻ 512 pixels with a pixel size of 0.391 mm for 56 cases, and a pixel size of 0.410 mm for four cases. The 3D MRA images included 128 slices for 55 cases with a slice thickness of 0.5 mm, 96 slices for two cases with a slice thickness of 0.67 mm, and 64 slices for three cases with a slice thickness of 1.0 mm. All original 3D MRA images were converted to isotropic volume data that were used for training and testing in this study by use of linear interpolation and/or cropping, where each of the volume data was 400 ⫻ 400 ⫻ 128 voxels with a voxel size of 0.5 mm. The clinical cases used in this study consisted of 29 cases with 36 aneurysms (diameter measured by radiologists, 3 to 26 mm; mean, 6.6 mm) and 31 non-aneurysm cases. The 31 non-aneurysm cases included 26 normals and 5 abnormal cases with other vascular diseases (ie, old brain infarction, old brain hemorrhage, intracranial steno-occlusive disease, meningioma, pituitary microadenoma, and azygous anterior cerebral artery), whereas 12 of the 29 aneurysm cases also included other vascular diseases. Figure 1 shows the distributions of measured diameters for the 36 unruptured aneurysms. Approximately two thirds of all aneurysms were smaller than 6.0 mm, and one aneurysm was very large. Thirty-four aneurysms were saccular in shape, and two were fusiform. Figure 2 shows the distribution of aneurysms at various locations, which was considered similar to the distribution in a clinical environ-
Academic Radiology, Vol 11, No 10, October 2004
DETECTING UNRUPTURED ANEURYSMS IN 3D MRA
Figure 2. Distribution of 36 aneurysms at various locations, including internal carotid artery (ICA) and middle cerebral artery (MCA), anterior communicating artery (ACoA), basilar artery (BA), anterior cerebral artery (ACA), posterior cerebral artery (PCA), and vertebral artery (VA).
ment. Approximately two thirds of all aneurysms were found on the internal carotid artery and middle cerebral artery. Overall Scheme Figure 3 shows the overall CAD scheme developed in this study. First, the isotropic 3D MRA images were processed by use of three selective filters (16) for enhancement of aneurysms, vessels, and vessel walls. The initial candidates were identified by use of a multiple gray-level thresholding technique on the 3D dot-enhanced images within the search area, which was determined by dilation of major vessels (17), because most aneurysms appear on specific vessels, as shown in Figure 2. Candidate regions were segmented by use of a region-growing technique with monitoring some image features. In the next step, all candidates were classified into four types according to their size and local structures. In each group, a number of false positives were removed by use of rules based on localized image features related to gray levels and morphology. Finally, linear discriminant analysis was used for further removal of false positives. Selective Enhancement Filters Recently, Li et al (16) developed three selective enhancement filters based on the eigenvalues of a Hessian matrix with multiscales for dot, line, and plane, which can simultaneously enhance objects of specific shapes (such as dot-like lung nodules) and suppress objects of other shapes (such as line-like vessels). Because some aneurysms are round protrusions and others are bal-
Figure 3. Computerized scheme for automated detection of unruptured intracranial aneurysms in 3D MRA.
loon-like objects, which appear on intracranial vessels, many aneurysm shapes were hemispherical or spherical. Therefore, for enhancement of aneurysms and suppression of other objects such as vessels, the isotropic 3D MRA images were processed by use of the dotenhancement filter, and the dot-enhanced images were used for identification of initial aneurysm candidates and segmentation of candidate regions. Furthermore, the isotropic 3D images were processed by use of lineand plane-enhancement filters for “vessels” (ie, blood flow), and vessel walls (ie, surface of blood flow) for determination of localized image features (average voxel value and standard deviation [SD] of voxel value) of aneurysm candidates, because these image features for aneurysms would be different from those for non-aneurysms in the line- and plane-enhanced images. Figure 4 shows an original MRA image and three images selectively enhanced for dot, line, and plane objects, all of which were produced by maximum intensity projection image processing. In the dot-enhanced image (Fig 4b), an aneurysm was enhanced well and “vessels” disappeared, although some “nonaneurysms” were also enhanced, which included bending regions and vessel bifurcations. On the other hand, the aneurysm disappeared in the line-enhanced image
1095
ARIMURA ET AL
Figure 4. Illustration of (a) an original MRA image and three images selectively enhanced for (b) dot, (c) line, and (d) plane objects, all of which were produced by maximum intensity projection image processing. Arrows indicate a large (7.5 mm) aneurysm.
(Fig 4c), but most of the vessels remained, and the walls of the aneurysm and vessels were enhanced in the plane-enhanced image (Fig 4d). Determination of Initial Aneurysm Candidates Identification of Initial Candidates Based on Multiple Gray-Level Thresholding For identification of initial aneurysm candidates, a multiple gray-level thresholding technique was applied to the dot-enhanced image that was smoothed by averaging with a square kernel (3 ⫻ 3 ⫻ 3) to reduce noise. Each threshold level was determined according to a certain spe-
1096
Academic Radiology, Vol 11, No 10, October 2004
cific upper percentage of the area under the pixel-value histogram in the dot-enhanced image within the search area. The pixel values of aneurysms in the dot-enhanced image were usually located at the high end of its histogram (the portion of the histogram closest to the highest pixel value), which ranged from approximately 0.008% to 0.8% for the cases used in this study. The regions in the dot-enhanced image above a certain threshold value were called “islands” (3D objects in 3D space), which were initial candidates. Note that the area under the histogram for each threshold level is equivalent to the total volume of all islands emerged. For initially picking up as many aneurysms as possible, the island volume and corresponding threshold level should be increased by a small enough volume for detecting a small aneurysm. Therefore, the incremental percentage of the area under the histogram for each threshold level was determined empirically by a fraction of a small volume relative to the total volume of the search area, which is equivalent to the total area of the histogram (eg, a small volume of approximately 20 mm3 and 85 mm3; the corresponding cube size of approximately 2.7 mm and 4.4 mm) for higher and lower threshold levels, respectively. At the first percentage threshold level where each island emerged (referred to as “starting percentage threshold level”), the effective diameter was determined for selection of initial aneurysm candidates. The effective diameter of a candidate was defined by the diameter of a sphere with the same volume as that of the candidate. If the effective diameter of an island was greater than 2 mm at the starting percentage threshold level, the island was considered an initial aneurysm candidate. Thus, an initial candidate selected at a starting percentage threshold level would not be examined again at the subsequent percentage threshold levels. Determination of Candidate Regions by Use of RegionGrowing Technique on Dot-Enhanced Image For each of the initial candidates, its region was determined by application of a region-growing technique to the dot-enhanced image to obtain the image features of the candidates for subsequent rule-based schemes. The candidate regions were determined within a volume of interest (40 mm ⫻ 40 mm ⫻ 40 mm), where the center of the volume of interest was located at the voxel with a maximum value for each initial candidate in the dot-enhanced image. Our segmentation method of the candidate region was based on finding a large change (referred to as a
Academic Radiology, Vol 11, No 10, October 2004
“transition point”) in some image features, which implied that the candidate region merged with its adjacent background structures or other candidates, as the candidate region grew. The region growing began at the location where the voxel value was the maximum in the initial candidate region, and was repeated at various gray levels, which were decreased from each previous gray level with a decrement of 5% of the maximum voxel value. The percentage of a gray-level decrease from the maximum value, which was referred to as “percentage gray level,” changed from 5% to 90%. At each percentage gray level, in this study, two image features for the candidate grown region (ie, the effective diameter and average contrast) were determined for finding the transition point in these features. The average contrast was defined by the difference in the average voxel values between the inside and outside regions in the original image divided by the average voxel value of the inside region. The inside and outside regions for each candidate were defined by the candidate region at a current percentage gray level and the increased region at the subsequent percentage gray level, respectively. However, for candidates without transition points in the effective diameter and the average contrast, the candidate regions were determined at the 90% gray level. Localized Feature Analysis and False Positive Removal Based on Rules According to Size and Local Structures All initial candidates were grouped into small and large candidates based on their effective diameters, some of the small and large false positives were then removed by use of different rules in each group (the first rulebased scheme). The remaining small candidates were further classified into three groups according to the local structures based on the skeleton image, which included a short-branch type, a single-vessel type, and a bifurcation type (including trifurcation), and some of the three types of false positives were removed by use of another set of different rules in each group (the second rule-based scheme). Classification of Initial Candidates Based on Effective Diameter It is important to note that the characteristics of large aneurysms in MRA images were quite different from those of small aneurysms. For example, the voxel values in the core region (near the center) for large aneurysms were commonly lower than those in the rind region adjacent to the
DETECTING UNRUPTURED ANEURYSMS IN 3D MRA
aneurysm wall because of the slow speed or turbulence of blood flow inside the aneurysm. By observing the original axial images of many aneurysms, we found that aneurysms with diameters larger than approximately 6.5 mm had this unique characteristic. In this study, initial candidates were grouped into small and large candidates by use of an effective diameter of 6.5 mm. First Rule-Based Schemes for Removal of Small and Large False Positives For the initial removal of small and large false positives, we used the image features of the gray level, size, and shape. Generally, the sizes of some false positives were smaller or larger than those of aneurysms, and some were less circular or more irregular compared with aneurysms. The average voxel values, SDs, and contrasts of some false positives were smaller or larger than those of aneurysms because of slow blood flow in some small vessels or fast speed in the bifurcation or bending regions on some vessels, respectively, and also for some other reasons caused by the turbulence of the blood flow inside some aneurysms and the nonuniform blood flow speed inside some vessels with nonuniform diameters. Therefore, we determined the gray-level features in the dot-enhanced and original images, ie, the average voxel value, the relative SD of the voxel value, the relative contrast, the average contrast (defined previously), the relative difference in the SD of voxel values between the candidate and outside regions, and the morphologic features, ie, the effective diameter, the sphericity, the relative SD of the distance between the centroid and the surface, and the maximum and minimum distance between the centroid and the surface. The “relative value” means the value relative to the average value. The degree of sphericity was defined by the fraction of the overlap volume between the candidate region and the sphere (with the same volume as the candidate volume). The relative SDs of voxel values were obtained in both the candidate and outside regions because local structures inside and outside aneurysms could differ from those for false positives. The relative contrast used in this study was defined by the difference between the maximum and minimum voxel values within the candidate region divided by the average voxel value. The average voxel values and the relative SD of the voxel values were determined in the line- and plane-enhanced images as well, because, as shown in Figure 4, aneurysms almost disappeared, but vessels were enhanced well in the line-enhanced images, and the appearances inside the aneurysm surfaces were different
1097
ARIMURA ET AL
from those of the non-aneurysm vessels in the plane-enhanced images. The relative SD of the distance between the centroid and the surface was related to the degree of irregularity for the surface of the candidate region. In addition, for removal of large false-positives, we determined specific features relevant to the characteristics of large aneurysms, ie, the average voxel values in the core and rind regions on the original image, the relative SDs of the voxel values in both regions, the relative differences in the average voxel values, and the SDs between the core and rind regions. All of the rules used in the two rule-based schemes were based on removal of false positives by use of simple thresholding for both the upper and lower limits of the features determined from all aneurysms included in each group; upper and lower limits were obtained, respectively, as 5% higher and 5% lower than the maximum and minimum values of each feature. If one of the features for a candidate was larger than the upper limit or smaller than the lower limit, the candidate was removed as a false positive. Classification of Small Candidates According to Local Structures Based on Skeleton Images It should be noted that the image features for small aneurysms and non-aneurysms were different in each group because of differences in the types of local structures. In addition, candidates with the short branch in a skeleton image (referred to as “short-branch type”) should be examined carefully because this short branch could be an aneurysm in an original image; details are explained in the Appendix. Therefore, we classified small candidates into three groups (short-branch type, single-vessel type, and bifurcation type), and established a number of effective rules to remove many non-aneurysms in each group. The grouping was made according to the local structures based on the skeleton image, which was obtained from the distance-transformed image, because the topologic properties (based on connectivity) would be preserved in the skeleton image (18,19). For example, the structure of a vessel would be simplified by maintaining voxels with only 1 voxel width. The local structure of the candidate was determined by counting the number of the skeleton objects in a rind region (a rind thickness of 1.0 mm) of a sphere of diameter 1.4 times larger than the candidate region. For example, if the number of skeleton objects was one or two, the local structure would be a vessel end or a single vessel, respectively. The details on the method
1098
Academic Radiology, Vol 11, No 10, October 2004
for classification of small candidates are described in the Appendix. Second Rule-Based Schemes for Removal of Small FalsePositives In the second rule-based scheme for further removal of many small false-positives (more than 80% false positives) for each group of the candidates, we established respective rules based on the localized image features. In this rule-based scheme, additional localized features were determined in both the distance-transformed image and the candidate region segmented in the original image. Note that the distance value in the distance-transformed image was related to the thickness of the vessel or the diameter of the vessel cross-section. In general, because small aneurysms may be considered as protrusions on the vessels, the diameter of the vessel cross-section with the aneurysm would be larger than the vessel without the aneurysm. Therefore, the maximum distance values in the candidate and outside regions, and the average distance values in the outside regions were obtained from the distance-transformed image, and also the relative difference in the maximum distance values (and the average distance values) between the candidate and the outside regions were determined. For the segmentation of the candidate region in the original image, we applied a region-growing technique as previously described. However, we found that not only the regions of the aneurysms (and false-positives) but also connected vessels were segmented at this stage. Therefore, the candidate region in the original image was segmented by region-growing only within the dilated volume of the candidate region obtained in the dot-enhanced image. However, for candidates of the short-branch type, the dilated volume was derived by dilation of the skeleton image of the short branch, because the candidates of the short-branch type were found based on the skeleton image, not the dot-enhanced image. The morphologic features for the candidate region in the original image included the effective diameter, the sphericity, the relative SD of the distance between the centroid and the surface, and the maximum and minimum distances between the centroid and the surface, which were the same as those used for the candidate region obtained in the dot-enhanced image. Additional features were determined for the short-branch type candidates (ie, the protrusion length and the average distance value in the candidate region obtained from the distance-transformed image). All of the rules used in the second rule-based scheme
Academic Radiology, Vol 11, No 10, October 2004
DETECTING UNRUPTURED ANEURYSMS IN 3D MRA
were based on removal of false positives by use of the same simple thresholding (upper and lower limits) as those used in the first rule-based scheme described previously. Linear Discriminant Analysis In our scheme, 43 image features can be used for the linear discriminant analysis for further removal of false positives. However, because the short-branch type candidates were detected only in the second step for the removal of small false positives, the number of features for the short-branch type candidates were limited (ie, the average voxel value, the relative SD of the voxel value, the relative contrast in the original image, and the effective diameter, the sphericity, the relative SD of the distance between the centroid and the surface, and the maximum and minimum distances between the centroid and the surface). Because some of the features were not useful for classification of the remaining candidates as aneurysms or false positives by use of a linear discriminant function, we selected the most effective combination of image features based on Wilks’ lambda and on the Az value (ie, the area under the receiver operating characteristic [ROC] curve) (20,21). Consequently, the final combination consisted of four features (ie, the average voxel value, the relative SD of the voxel value, the relative SD of the distance between the centroid and the surface, and the difference between the maximum and minimum distance [between the centroid and the surface]). Evaluation of the Performance Our CAD scheme was evaluated with both a consistency test and a leave-one-out-by-patient test method. With the consistency test, cases used for training were also used for testing. With the leave-one-out-by-patient test method, all candidates except one and candidates obtained from the same patient were used for training, and the one candidate left out was used for testing with two rule-based schemes and the linear discriminant function. This procedure was repeated for all candidates, so that each candidate was used once as a test candidate. By changing a threshold value for discriminant scores of the candidates produced with the linear discriminant function, which was determined by linear discriminant analysis for distinction between aneurysms and false positives, we determined the free-response receiver operating characteristic (FROC) curve of the CAD scheme in the leave-oneout-by-patient test method.
Figure 5. Results obtained with the segmentation of (a) small and (b) large aneurysms by use of the region-growing technique on the dot-enhanced or original images.
RESULTS AND DISCUSSION The computerized scheme for detection of the intracranial aneurysms in MRA was applied to 31 non-aneurysm cases and 29 abnormal cases with 36 aneurysms. All of the 36 aneurysms with 22.3 false positives per patient were detected at the initial identification step. For evaluation of our scheme in the computerized detection of aneurysms, we used a criterion such that an aneurysm was considered correctly detected if the location of the maximum voxel value in the candidate region was within the diameter of the aneurysm measured by radiologists for aneurysms smaller than 7.0 mm, and within the diameter of 7.0 mm for aneurysms larger than 7.0 mm. Figure 5 shows the results obtained with the segmentation of small and large aneurysms by use of the region-
1099
ARIMURA ET AL
Academic Radiology, Vol 11, No 10, October 2004
Figure 6. Relationship between the effective diameter in the dot-enhanced image and the relative SD of the voxel value in the original image for small and large aneurysms (circles and squares) as well as small and large false positives (dots and triangles) in the dot-enhanced image. Rules are indicated by dashed lines.
Figure 7. Relationship between the maximum distance value in the candidate region and average distance value in the outside region, for single-vessel-type candidates (aneurysms: circles; false positives: dots). Rules are indicated by dashed lines.
growing technique on the dot-enhanced or original images, where most of the small and large aneurysms were segmented well. However, we found that segmentation of large aneurysms tended to be less accurate; not only the aneurysm, but also adjacent vessel regions (or background) were included in the segmented candidate region, (eg, large aneurysms in the middle and bottom of Figure 5b). These inaccuracies of segmentation occurred because the contrasts of the surface area of such large aneurysms were very low, and the average voxel values in the core region for the aneurysms were lower than those in the rind region. Nevertheless, it was not difficult to distinguish between large aneurysms and large false positives because of the unique image features of large aneurysms. In the first rule-based scheme, all initial candidates were grouped into small and large candidates, and many false positives were removed by use of rules based on the localized image features. Figure 6 shows the relationship between the effective diameter in the dot-enhanced image and the relative SD of the voxel value in the original image. The effective diameters of some false positives, such as small short vessels and large long vessels, were smaller and larger, respectively, than those of aneurysms. In addition, because the distribution of voxel values inside some vessels (eg, vessels of elderly patients with
stenosis and occlusion) were more nonuniform than those of aneurysms, the relative SDs of the voxel values for these false positives were greater than those of aneurysms. Therefore, many false positives were removed by use of such rules as shown by dashed lines in Figure 6. By use of the first rule-based scheme, all aneurysms were retained, and the average number of false positives per patient was reduced from 22.3 to 5.8. At this stage, the majority of the remaining false positives were due to bending of single vessels and bifurcation vessels, most of which were further removed by use of the second rulebased scheme, as described below. For single-vessel-type candidates, Figure 7 shows the relationship between the average distance values in the outside region and the maximum distance values in the candidate region, where the distance values were obtained from the distance-transformed images. Note that many bending regions of large vessels were included as falsepositives in single-vessel-type candidates, and the average distance values in the outside region for such false-positives were larger than those of aneurysms, as shown by many false positives on the right side in Figure 7. In addition, the difference in the distance values between the candidate region and the outside region tended to be larger for aneurysms than those for most of the false posi-
1100
Academic Radiology, Vol 11, No 10, October 2004
Figure 8. FROC curve for overall performance of our computerized scheme in automated detection of intracranial aneurysms in MRA images.
tives. Finally, many false positives in each group were removed based on the differences in the image features between aneurysms and false positives by use of rules as shown by dashed lines in Figure 7. As a result, all of the 36 aneurysms were detected correctly with 0.55 falsepositives per patient in a consistency test. Furthermore, as a result of the leave-one-out-by-patient test method, our scheme achieved a sensitivity of 100% with 2.4 falsepositives per patient. Figure 8 shows the FROC curve for the overall performance of our scheme by use of linear discriminant analysis with the leave-one-out-by-patient test method. According to this result, our CAD scheme had a sensitivity of 89% with 1.3 false-positives per patient. At this operating point, four aneurysms were not detected by our CAD scheme, and thus may be considered difficult cases for a computerized detection. These cases were two large aneurysms and two short-branch type small aneurysms. Segmentation of the large aneurysms could be inaccurate as shown in Figure 5, because the contrasts of the surface area of large aneurysms were very low and the average voxel values in the core region for large aneurysms were lower than those in the rind region. The segmented regions of two large aneurysms included adjacent vessels and some cavities. Short-branch type small aneurysms adjacent to large regions, especially the bending region of a large parent vessel, were hardly enhanced in the dot-
DETECTING UNRUPTURED ANEURYSMS IN 3D MRA
enhanced image, because such large regions tended to be strongly enhanced by the dot-enhancement filter. Although some of small aneurysms were detected by finding the short branches, the segmentations for some of shortbranch type aneurysms tended to be relatively inaccurate. Therefore, we need to improve the segmentation of large aneurysms and short-branch type aneurysms in the future. During the last two decades, we developed a number of CAD schemes for detection and classification of various abnormalities such as microcalcifications and masses in mammograms (22,23), pulmonary nodules and interstitial infiltrates in chest radiographs (24,25), and nodules and diffuse lung diseases in computed tomography (26,27). We have shown the usefulness of these schemes by carrying out a number of observer performance studies (28,29). Therefore, for investigating the usefulness of our CAD scheme for detection of intracranial aneurysms in MRA, it will be also necessary to carry out an observer study by use of ROC analysis to compare a radiologists’ performance without and with the aid of CAD output in terms of the detection accuracy of aneurysms in MRA. Our database included many different kinds of cases with aneurysms that were selected by considering the size, shape, and location of aneurysms, the gender and age of the patients, and other vascular diseases such as stenosis or occlusion. However, because all cases used in this study were obtained from a magnetic resonance imaging scanner in one hospital, our results would depend on the magnetic resonance imaging scanner and the acquisition sequence used in the study. Furthermore, the number of the false positives or specificity would depend on the patients with other vascular diseases, because the appearance of the blood flow within such vessels in the MRA images might look like some aneurysms for the computer. Therefore, our CAD scheme needs to be applied to an independent database including many cases acquired from different magnetic resonance imaging scanners on many different patients. In addition, because all rules used in this study were determined based on image features of a relatively small number of aneurysms (ie, from 6 to 11 in each group), the number of abnormal cases needs to be increased for determination of more general and robust rules for unknown cases. It has been reported that screening of intracranial aneurysms would be especially useful in high-risk groups (eg, patients with adult polycystic kidney disease or a strong family history of aneurysm SAH) (30 –33). Magnetic resonance angiography may be used as screening examinations in detecting intracranial aneurysms, where our CAD
1101
ARIMURA ET AL
scheme could aid radiologists in detecting aneurysms; this may be similar to the usefulness of the CAD scheme in the detection of breast cancers in screening mammograms (34). In conclusion, we have developed a CAD scheme based on the use of selective enhancement filters for the automated detection of unruptured intracranial aneurysms in MRA. With our CAD scheme, all of the 36 aneurysms in 60 cases were detected correctly with 0.55 false-positives per patient in an evaluation by a consistency test, and 2.4 false-positives by the leave-one-out-by-patient test method. Our CAD system would be useful in assisting radiologists in the detection of unruptured intracranial aneurysms in MRA.
Academic Radiology, Vol 11, No 10, October 2004
14. 15.
16.
17.
18.
19.
ACKNOWLEDGMENT
The authors thank Junji Shiraishi, PhD, Kenji Suzuki, PhD, Feng Li, MD, Chisako Muramatsu, BS, and Roger Engelmann, MS, for their useful discussions, and Elisabeth Lanzl for improving the manuscript.
20. 21.
22.
REFERENCES 1. Wardlaw JM, White PM. The detection and management of unruptured intracranial aneurysms. Brain 2000; 123:205–221. 2. Weir B. Unruptured intracranial aneurysms: A review. J Neurosurg 2002; 96:3– 42. 3. Fogelholm R, Hernesniemi J, Vapalahti M. Impact of early surgery on outcome after aneurysm subarachnoid hemorrhage: a populationbased study. Stroke 1993; 24:1649 –1654. 4. Hop JW, Rinkel GJ, Algra A, van Gijn J. Case-fatality rates and functional outcome after subarachnoid hemorrhage: A systematic review [review]. Stroke 1997; 28:660 – 664. 5. Hijdra A, Braakman R, van Gijn J, Vermeulen M, van Crevel H. Aneurysmal subarachnoid hemorrhage: Complications and outcome in a hospital population. Stroke 1987; 18:1061–1067. 6. Sengupta RP, McAllister VL. Subarachnoid hemorrhage. Berlin: SpringerVerlag, 1986. 7. Broderick JP, Brott TG, Duldner JE, Tomsick T, Leach A. Initial and recurrent bleeding are the major causes of death following subarachnoid hemorrhage. Stroke 1994; 25:1342–1347. 8. Juvela S, Porras M, Poussa K. Natural history of unruptured intracranial aneurysms: Probability of and risk factors for aneurysm rupture. J Neurosurg 2000; 93:379 –387. 9. Ogawa T, Okudera T, Noguchi K, Sasaki N, Inugami A, Uemura K, Yasui N. Cerebral aneurysms : Evaluation with three-dimensional CT angiography. AJNR Am J Neuroradiol 1996; 17:447– 454. 10. Korogi Y, Takahashi M, Katada K, Ogura Y, Hasuo K, Ochi M, Utsunomiya H, Abe T, Imakita S. Intracranial aneurysms: Detection with three-dimensional CT angiography with volume rendering – Comparison with conventional angiographic and surgical findings. Radiology 1999; 211:497–506. 11. White PM, Teasdale EM, Wardlaw JM, Easton V. Intracranial aneurysms: CT angiography and MR angiography for detection – Prospective blinded comparison in a large patient cohort. Radiology 2001; 219:739 –749. 12. Preda L, Gaetani P, Rodriguez y Baena R, et al. Spiral CT angiography and surgical corrections in the evaluation of intracranial aneurysms. Eur Radiol 1998; 8:739 –745. 13. Huston J III, Nichols DA, Luetmer PH, Goodwin JT, Meyer FB, Wiebers DO, Weaver AL. Blinded prospective evaluation of sensitivity of MR
1102
23. 24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
angiography to known intracranial aneurysms: Importance of aneurysm size. AJNR Am J Neuroradiol 1994; 15:1607–1614. Aprile I. Evaluation of cerebral aneurysms with MR-angiography. Riv Neuroradiol 1996; 9:541–550. Korogi Y, Takahashi M, Mabuchi N, et al. Intracranial aneurysms: Diagnostic accuracy of three-dimensional, Fourier transform, time-o-flight MR angiography. Radiology 1994; 193:181–186. Li Q, Sone S, Doi K. Selective enhancement filters for nodules, vessels, and airway wall in two- and three-dimensional CT scans. Med Phys 2003; 30:2040 –2051. Arimura H, Li Q, Korogi Y, et al. Development of CAD scheme for automated detection of intracranial aneurysms in magnetic resonance angiography. 18th International Congress of Computer Assisted Radiology and Surgery. Chicago, IL; International Congress Series 2004; 1268:1015–1020. Saito T, Toriwaki J. New algorithms for Euclidean distance transformations of an n-dimensional digitized picture with applications. Pattern Recognition 1994; 27:1551–1565. Saito T, Toriwaki J. A sequential thinning algorithm for three dimensional digital pictures using the Euclidean distance transformation. Proceedings of the 9th Scandinavian Conference on Image Analysis SCIA 95. Uppsala, Sweden 1995; 507–516. Johnson RA, Wichern DW. Applied Multivariate Statistical Analysis. Cliffs, NJ; Prentice-Hall, 1992, 5.3:184 –188. Aoyama M, Li Q, Katsuragawa S, et al. Automated computerized scheme for distinction between benign and malignant solitary pulmonary nodules on chest images. Med Phys 2002; 29:701–708. Chan HP, Doi K, Galhotra S, Vyborny CJ, MacMahon H, Jokich PM. Image feature analysis and computer-aided diagnosis in digital radiography. 1. Automated detection of microcalcifications in mammography Med Phys 1987; 14:538 –548. Kupinski MA, Giger ML. Automated seeded lesion segmentation on digital mammograms. IEEE Trans Med Imaging 1998; 17:510 –517. Giger ML, Doi K, MacMahon H. Image feature analysis and computeraided diagnosis in digital radiography. 3. Automated detection of nodules in peripheral lung fields Med Phys 1988; 15:158 –166. Katsuragawa S, Doi K, MacMahon H. Image feature analysis and computer-aided diagnosis in digital radiography. Detection and characterization of interstitial lung disease in digital chest radiographs Med Phys 1988; 15:311–319. Suzuki K, Armato SG III, Li F, Sone S, Doi K. Massive training artificial neural network (MTANN) for reduction of false positives in computerized detection of lung nodules in low-dose CT. Med Phys 2003; 30: 1602–1617. Uchiyama Y, Katsuragawa S, Abe H, et al. Quantitative computerized analysis of diffuse lung disease in high-resolution computed tomography. Med Phys 2003; 30:2440 –2464. Chan HP, Doi K, Vyborny CJ, et al. Improvement in radiologists’ detection of clustered microcalcifications on mammograms – The potential of computer-aided diagnosis. Invest Radiol 1990; 25:1102–1110. Shiraishi J, Abe H, Engelmann R, Aoyama M, MacMahon H, Doi K. Computer-aided diagnosis for distinction between benign and malignant solitary pulmonary nodules in chest radiographs: ROC analysis of radiologists’ performance. Radiology 2003; 227:469 – 474. Levey AS. Screening for occult intracranial aneurysms in polycystic kidney disease: Interim guidelines [editorial]. J Am Soc Nephrol 1990; 1:9 –12. Wiebers DO, Torres VE. Screening for unruptured intracranial aneurysms in autosomal dominant polycystic kidney disease [editorial]. N Engl J Med 1992; 327:953–955. Ronkainen A, Puranen MI, Hernesniemi JA, et al. Intracranial aneurysms: MR angiographic screening in 400 asymptomatic individuals with increased familial risk. Radiology 1995; 195:35– 40. Kojima M, Nagasawa S, Lee YE, Takeichi Y, Tsuda E, Mabuchi N. Asymptomatic familial cerebral aneurysms. Neurosurgery 1998; 43:776 –781. Freer TW, Ulissey MJ. Screening mammography with computer-aided detection: Prospective study of 12,860 patients in a community breast center. Radiology 2001; 220:781–786.
Academic Radiology, Vol 11, No 10, October 2004
Figure A1. Schematic diagram for classification of small candidates into three groups including short-branch type, singlevessel type, and bifurcation type based on the skeleton image.
APPENDIX The schematic diagram for classification of small candidates into three groups based on the skeleton image is shown in Figure A1. First, the candidate region was segmented within the volume of interest (30 mm ⫻ 30 mm ⫻ 30 mm) in the original image, where the centroid of the candidate region was located at the center of the volume of interest by use of the region-growing technique, which is similar to that used for the dotenhanced image. The distance-transformed image was derived by calculation of a Euclidean minimum distance from each voxel in the segmented region to the nearest background in the binary image (18). Next, the skeleton image was obtained by use of a thinning algorithm (19) based on the distance-transformed image, where deletable voxels with smaller distance values were removed first so that the topologic properties of the segmented vessels could be preserved. For classification of the local structures, the skeleton images of candidates were analyzed as shown in Figure A1. Note that “lump” candidates and short-branch type candidates were identified in this classification scheme, as described in detail in the next two paragraphs. If the degree of lump for the candidate (defined in the next paragraph) was smaller than 0.1, the number of skele-
DETECTING UNRUPTURED ANEURYSMS IN 3D MRA
ton objects was counted in the rind region. If not, candidates were classified as the bifurcation type, because most of the aneurysms with a degree of lump greater than 0.1 were located at the bifurcation. If the number of skeleton objects in the rind region was 1, the local structure would be considered as a single vessel. If the number of the skeleton objects was ⱖ2, the nearest short branch was searched in the large sphere including the original candidate. If a short branch was found, then the candidate was determined as the short-branch type. If the short branch was not found, and if the number of the skeleton objects was equal to 2, the candidate would be a single-vessel type; if the number of skeleton objects was ⱖ3, the candidate would be a bifurcation type. The skeleton objects for some candidates would have a “lump” composed of many short “skeletons,” which look like short hairs and would not be true skeletons; therefore, the number of the skeleton objects in the rind region could be incorrect. Thus, before counting the skeleton objects in the rind region, we determined the degree of lump for each candidate for identifying the lump candidates, where the degree of lump was defined by a fraction of the total volume of skeleton objects in a 2.0-mm sphere placed at the centroid of the candidate region. A small protrusion, a small branching vessel (nonaneurysm), and a small aneurysm on a single vessel or bifurcation (parent vessel) in the original image may be considered as a short branch attached to a parent skeleton in the skeleton image, as shown in Figure A2. Such candidates with a short branch were referred to as short-branch type and should be examined carefully because this short branch could be an aneurysm. Note that some small aneurysms adjacent to large regions, such as the bending region of a large parent vessel, were hardly enhanced in the dot-enhanced image because such large regions tended to be strongly enhanced by the dot-enhancement filter. Such small aneurysms could be detected by finding the short branch adjacent to the original candidate region in the skeleton image. Thus, the short branch was searched as illustrated in Figure A2: (1) the nearest short branch with an end point was searched within a sphere of diameter four times larger than the original candidate region, and the large sphere was placed at the centroid of the original candidate region; (2) if the short-branch length was larger than the radius of a parent vessel, and if the difference between the length and the radius was in the
1103
ARIMURA ET AL
Academic Radiology, Vol 11, No 10, October 2004
range from 1.5 mm to 6.5 mm, then the short branch could be a small protrusion on the parent vessel or a small aneurysm; and thus, (3) the short branch was considered as a new candidate of the short-branch type at this stage. Otherwise, the candidate was classified as either the bifurcation type or the single-vessel type. The difference between the short-branch length and the radius of a parent vessel was defined as the protrusion length, which was obtained as an image feature for the short-branch type candidates used in the second rulebased scheme. Because both the short branch and the vessel end have an end point, it was difficult to distinguish between them. Therefore, we decided not to examine whether the vessel-end candidates were of the short-branch type, and the vessel-end candidates were classified as the single-vessel type in this study, because a vessel end would consist of a single vessel. Figure A2. Illustration of a method for searching for a short branch adjacent to the original candidate region in the skeleton image.
The Association of University Radiologists wishes to thank the following companies for their generous support of the AUR Junior Membership Program:
1104
4th Year Radiology Residents
3rd Year Radiology Residents
Siemens Medical Systems, Inc.
Berlex Laboratories