1076 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 22, NO. 9, SEPTEMBER 2003 Limits on the Accuracy of 3-D Thickness Measurement in Magnetic Resonance Images—Effects of Voxel Anisotropy Yoshinobu Sato*, Member, IEEE, Hisashi Tanaka, Takashi Nishii, Katsuyuki Nakanishi, Nobuhiko Sugano, Tetsuya Kubota, Hironobu Nakamura, Hideki Yoshikawa, Takahiro Ochi, and Shinichi Tamura, Member, IEEE Abstract—Measuring the thickness of sheet-like thin anatomical structures, such as articular cartilage and brain cortex, in threedimensional (3-D) magnetic resonance (MR) images is an important diagnostic procedure. This paper investigates the fundamental limits on the accuracy of thickness determination in MR images. We defined thickness here as the distance between the two sides of boundaries measured at the subvoxel resolution, which are the zero-crossings of the second directional derivatives combined with Gaussian blurring along the normal directions of the sheet surface. Based on MR imaging and computer postprocessing parameters, characteristics for the accuracy of thickness determination were derived by a theoretical simulation. We especially focused on the effects of voxel anisotropy in MR imaging with variable orientation of sheet-like structure. Improved and stable accuracy features were observed when the standard deviation of Gaussian blurring combined with thickness determination processes was around 2 2 times as large as the pixel size. The relation between voxel anisotropy in MR imaging and the range of sheet normal orientation within which acceptable accuracy is attainable was also clarified, based on the dependences of voxel anisotropy and the sheet normal orientation obtained by numerical simulations. Finally, in vitro experiments were conducted using an acrylic plate phantom and a resected femoral head to validate the results of theoretical simulation. The simulated thickness was demonstrated to be wellcorrelated with the actual in vitro thickness. Index Terms—Articular cartilage, brain cortex, point spread function, quantitative image analysis, spatial resolution, thickness determination, three-dimensional imaging. Manuscript received July 4, 2002; revised March 21, 2003. This work was supported in part by the Japan Society for the Promotion of Science (JSPS) Research for the Future Program JSPS-RFTF99I00903 and JSPS Grant-in-Aid for Scientific Research (B)(2)15300059. The Associate Editor responsible for coordinating the review of this paper and recommending its publication was J. Duncan. Asterisk indicates corresponding author. *Y. Sato is with the Division of Interdisciplinary Image Analysis, Osaka University Graduate School of Medicine, 2-2-D11, Yamada-oka, Suita, Osaka, 5650871, Japan (e-mail: [email protected]). H. Tanaka, K. Nakanishi, and H. Nakamura are with the Department of Radiology, Osaka University Graduate School of Medicine, Suita, Osaka, 565-0871, Japan. T. Nishii, N. Sugano, and H. Yoshikawa are with the Department of Orthopaedic Surgery, Osaka University Graduate School of Medicine, Suita, Osaka, 565-0871, Japan. T. Kubota is with Tsuyama National College of Technology, 624-1, Numa, Tsuyama, Okayama 708-8509, Japan. T. Ochi is with the Division of Applied Medical Engineering, Osaka University Graduate School of Medicine, Suita, Osaka, 565-0871, Japan. S. Tamura is with the Division of Interdisciplinary Image Analysis, Osaka University Graduate School of Medicine, Suita, Osaka, 565-0871, Japan. Digital Object Identifier 10.1109/TMI.2003.816955 I. INTRODUCTION T HICKNESS measurement of sheet-like or plate-like thin anatomical structures in magnetic resonance (MR) images is an important procedure in clinical practice. For example, diagnosis of joint diseases requires the evaluation for the distribution of articular cartilage thicknesses [1]–[8], and diagnosis of specific neuropsychiatric disorders needs the assessment of cortical thicknesses in the brain [9]–[11]. While several methods for thickness quantification have been proposed [5], [6], [8], [9], inherent limits on accuracy arising from finite resolution have not been adequately evaluated. Previously, Sato et al. examined the limits on accuracy resulting from partial volume averaging with emphasis on the effect of anisotropic voxels by software simulation [8]. However, the modeling of MR image acquisition was insufficient and validation with actual MR images was not carried out in that study. The aim of the present study is, therefore, to provide a theoretical procedure for ascertaining the inherent limits on the accuracy of thickness determination in MR images. A preliminary report was communicated earlier [12]. In the present paper, thickness is defined as the distance between the two sides of boundaries of sheet structures measured at the subvoxel resolution, which are the zero-crossings of the second directional derivatives along the normal directions of the sheet surface. We focus on the effects of the parameters relevant to computer postprocessing as well as those of MR imaging parameters on thickness accuracy. For the effects of computer postprocessing parameters, the dependences of the standard deviation (SD) in Gaussian blurring combined with thickness determination processes are examined, and its role in obtaining improved and stable accuracy characteristics is elucidated. To determine the effects of MR imaging parameters, we especially address the question as to how the accuracy depends on the orientation of sheet structures when a voxel shape is anisotropic. Sheet structures of interest are often distributed within specific orientation ranges. If they are situated on an approximated cylindrical surface, the voxel shape can be highly anisotropic to balance the signal-to-noise ratio (SNR) and thickness accuracy. As a result, the resolution in the plane orthogonal to the axis of the cylinder is much higher than that along the axis. However, if their orientations are randomly distributed or unknown, isotropic (cubic) voxel shape might be the best. We establish a method for evaluating the dependences of sheet orientation, voxel anisotropy, and Gaussian SD on the accuracy of thickness measurement 0278-0062/03$17.00 © 2003 IEEE SATO et al.: LIMITS ON THE ACCURACY OF 3-D THICKNESS MEASUREMENT IN MAGNETIC RESONANCE IMAGES by numerical simulations based on mathematical modeling of sheet structures, MR imaging, and postprocessing. The simulation method is validated by in vitro experiments with the actual measurements. The characteristics of accuracy determination obtained by the simulations are useful in delivering objective criteria for designing optimal imaging and postprocessing protocols, provided that the orientation distribution of sheet structures and the fixed volume of a voxel directly related to SNR are known. 1077 (a) II. THEORY A. Modeling a Sheet Structure A three-dimensional (3-D) sheet structure orthogonal to the axis is modeled as (1) where , and (b) (2) , and in which represents the thickness of the sheet. , are the MR signal intensities of the sheet and both sides of backgrounds, respectively [Fig. 1(a)]. Let ( , ) be a pair of latitude and longitude which represents the normal orientation of the sheet given by (3) The 3-D sheet structure with orientation Fig. 1. Modeling 3-D sheet structures. (a) Bar profile of MR values along sheet normal direction with thickness . L , L , and L denote sheet object, left-side, and right-side background levels, respectively. (b) Three-dimensional r . sheet structures with thickness and normal orientation ~ of the sheet structure with orientation given by and thickness is (9) where denotes the convolution operation. is written as (4) , in which denotes a 3 3 matrix reprewhere senting rotation which causes the normal orientation of the sheet , i.e., the axis, to correspond to [Fig. 1(b)]. B. Modeling MR Image Acquisition The one-dimensional (1-D) point spread function (PSF) of MR images [13] is given by (5) is the number of samples in the frequency domain, where represents the sampling interval in the spatial domain. and Equation (5) is well-approximated [14] by (6) C. Thickness Determination Procedure In this paper, we restrict the scope of our investigation to the sheet model described in Section II-A (Fig. 1), that is, a sheet . We destructure with constant thickness and orientation fine the thickness measured from the MR imaged sheet structure as the distance between both sides of image edges along the sheet normal vector. As long as the sheet model shown in Fig. 1 is considered, other definitions of measured thickness, for example, the shortest distance between both sides of the image edges, generally give the same thickness value. We define the image edges as the zero-crossings of the second directional derivatives along the sheet normal vector, which is equivalent to the Canny edge detector [16]. Gaussian blurring is typically combined with the second directional derivatives to adjust scale as well as reduce noise. The partial second derivative combined with Gaussian blur, for example, is given by ring for the MR image (10) where where (7) (11) The 3-D PSF is given by (8) , , and are sampling intervals along the axis, where axis, and axis, respectively. In actual MR imaging, the magnitude operator is applied to the complex number obtained at each voxel by FFT reconstruction, whose effects are not negligible [15]. Thus, the MR image is the isotropic 3-D Gaussian function in which is repwith SD . The second directional derivative along resented as (12) , in which denotes a 3 3 matrix reprewhere senting rotation which causes the normal orientation of the sheet 1078 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 22, NO. 9, SEPTEMBER 2003 , i.e., the axis, to correspond to larly, the first directional derivative along [Fig. 1(b)]. Simiis represented as (13) Practically, the second and the first directional derivatives can be calculated in a computationally efficient manner using the Hessian matrix and gradient vector, respectively. The second directional derivative along the normal direction of the sheet structure is given by (14) where is the Hessian matrix given by (15) Similarly, the first directional derivative along the sheet normal is given by (a) (16) where is the gradient vector given by (17) Thickness of sheet structures can be determined by analyzing and along straight 1-D profiles of line given by (18) is a parameter representing the position on the where and straight line. By substituting (18) for in , (b) (19) and (20) are derived, respectively. Fig. 2(a) shows a schematic diagram for the 1-D profile processing. Both sides of the boundaries for sheet structures can be defined as the points with the maximum among those satisfying the conand minimum values of . Let have its maximum and dition given by and , respectively. The meaminimum values at sured thickness, is defined as the distance between the two detected boundary points, which is given by (21) The procedures for thickness determination from discrete real 3-D MR images are described later in Section III-B3. D. Frequency Domain Analysis of MR Imaging and Thickness Determination In order to elucidate the effects of MR imaging and postprocessing parameters, observations in the frequency domain and are helpful. All the processes to obtain from the original sheet structure are modeled as linear filtering processes excepting the magnitude operator applied in (9). (c) Fig. 2. Thickness determination procedure using zero-crossings of the second directional derivatives along sheet normal direction. (a) Basic concept of thickness determination procedure. (b) Interpolation of discrete MR data. (c) Zero-crossing search procedure. 1) Modeling a Sheet Structure: The Fourier transform of , is given 3-D sheet structure orthogonal to the axis, by (22) denotes the unit where represents the Fourier transform, . Note that impulse, and when and in . The Fourier transform of 3-D sheet structure whose normal is , , is given by (23) SATO et al.: LIMITS ON THE ACCURACY OF 3-D THICKNESS MEASUREMENT IN MAGNETIC RESONANCE IMAGES 1079 where , in which denotes a 3 3 matrix repaxis correspond to resenting rotation which enables the [Fig. 3(a)]. has In the 3-D space of the frequency domain, energy only in the 1-D subspace represented as a straight line given by (24) where is a parameter representing the position on the straight line. By substituting (24) for in (23), the following is derived: (25) represents energy distribution along (24). Analysis where , is sufficient to exof the degradation of 1-D distribution, amine the effects of MR imaging and postprocessing parameters is the in the subsequent processes. It should be noted that . 1-D sinc function when 2) Modeling MR Image Acquisition: The Fourier transform of MR PSF is given by (a) (26) where [Fig. 3(b)], and otherwise. (27) By substituting (24) for in (26) to obtain 1-D frequency , the following is derived component affecting (b) (28) Thus, the Fourier transform of MR image of the sheet structure, is given by (29) represents the inverse Fourier transform. If is a nonnegative function, is and all the processes can given by be described as linear filtering processes. Deformation of the original signal due to truncation is clearly understandable in the frequency domain [Fig. 3(c)]. 3) Gaussian Derivatives of MR Imaged Sheet Structure: The Fourier transform of the second derivative of Gaussian of is given by where (30) and that of the second directional derivative along sented as is repre- (31) , in which denotes a 3 3 matrix where axis correspond to representing rotation which enables the . The 1–D frequency component of affecting is given by (32) (c) Fig. 3. Frequency domain analysis of sheet structure modeling, MR imaging, and thickness determination. (a) Modeling a sheet structure. In the frequency domain, a sheet structure is basically modeled as the sinc function whose width is inversely proportional to the thickness in the spatial domain. (b) Modeling MR imaging. It is assumed here that . The voxel size determines the frequency bandwidth of each axis, which is also inversely proportional to the size in the spatial domain. (c) Modeling MR image acquisition of a sheet structure. In the frequency domain, imaged sheet structure is essentially the band-limited sinc function. 1 =1 =1 Similarly, the 1-D component of the first directional derivative , is obtained. of Gaussian, 1080 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 22, NO. 9, SEPTEMBER 2003 Finally, the Fourier transforms of and given by and are derived (33) Hence, it can be assumed that the resolution along the axis is lower than that in the - plane and that pixels in the - plane , and a measure of voxel anisotropy are square, i.e., . In the simulations, we assumed that can be defined as (35) and (34) respectively. The 1-D profiles along the sheet normal direction of the Gaussian derivatives of MR imaged sheet structures [(19) and (20)] are obtained by inverse Fourier transform of (33) and (34), and then thickness is determined according to the procedure shown in Fig. 2(a). While simulating the MR imaging and Gaussian derivative computation described in Section II-C essentially requires 3-D convolution in the spatial domain, only 1-D computation is necessary in the frequency domain, which drastically reduces computational cost. In the following sections, we examine the effects of various parameters, which are involved in the sheet model, MR imaging resolution, and thickness determination processes, on measurement accuracy. Efficient computational methods of simulating MR imaging and postprocessing thickness determination processes are essential, and thus simulating the processes by 1-D signal processing in the frequency domain is regarded as the key to comprehensive analysis. III. MATERIALS AND METHODS A. Numerical Simulation In order to examine the effects of various parameters on the accuracy of thickness determination, numerical simulation based on the theory described in Section II was performed. The parameters used in the simulation were classified into , , , and for the following three categories: , , , and for determining MR defining sheet structures, imaging resolution, and Gaussian SD, , used in computer postprocessing for thickness determination. was obWe assumed that the estimated sheet thickness tained under the condition that the sheet normal orientation was known. The numerical simulation was performed in the frequency domain exactly in the same manner as described in Sec, , , tion II-D. Based on sheet structure parameters , , MR imaging parameters , , and , and postand and were obtained by processing parameter , 1-D computation in the frequency domain according to (33) and and were obtained by (34), respectively. And then, and , respectively. inverse Fourier transform of and , thickness was estimated using (21). Using Finally, estimated thickness was compared with the actual thickness to reveal the limits on accuracy. It should be noted that only 1-D computation was necessary for 3-D thickness determination in our numerical simulation. In the simulation, the effect of anisotropic resolution of MR be the pixel volume data was the focus. Let size within the slices. Resolution of MR volume data is typically anisotropic because they usually have lower resolution along the third direction (orthogonal to the slice plane) than within slices. without loss of generality, and thus, (36) We performed the above described numerical simulation with , , , , , and . different combinations of , B. In Vitro Experiments To validate the numerical simulation, the postprocessing method for thickness determination was used to measure real MR images of two different objects, an acrylic plate phantom and a resected femoral head. 1) Plate Phantom: A phantom of sheet-like objects with known thickness was used. It consisted of four acrylic plates ) with thickness , 1.5, 2.0, and 3.0 of 80 80 ( (mm), placed parallel to each other with an interval of 30 mm [Fig. 4(a)]. The phantom was submerged in a water bath so that the background (water) showed higher intensity as contrasted to low intensity objects (acrylic plates). Three–dimensional MR images (TR/TE/flip angle/matrix/FOV/slice thickness: 12.8 ms/5.6 ms/5/256 256/160 mm/1.5 mm) of the phantom were obtained using a fast spoiled gradient-echo sequence (FSPGR). and The voxel size was . Thus (37) Thirteen data sets of 3-D MR images were acquired with different normal positions of the phantom plates, eight with variand ) and fixed able ( ( ), and five with variable ( and ) and fixed ( ). In the obtained MR images, we oband . Fig. 4(b) and (c) shows served examples of the MR images. We compared actually measured thickness from the real MR data with the computational thickness calculated by the numerical simulations. 2) Femoral Head Specimen: The resected femoral head used in the study was approximately spherical in shape, with articular cartilage distributed on its surface [Fig. 5(a)]. The cartilage was the subject of the experiments, in which we assumed it was distributed on a spherical surface. The articular cartilage was imaged as a bright sheet structure distributed on the spherical head surface. The thickness of the cartilage was then measured along the normal direction of the spherical surface approximating the femoral head. The cartilage did not completely satisfy the restrictions on the sheet model assumed in Section II, that is, constant thickness and orientation (Fig. 1). However, it did not contain large variations in its thickness and orientation and, thus, approximately satisfied the restrictions. 3-D MR images of sagittal sections were obtained using 3-D-spoiled gradient-echo sequences (SPGR) [7] under the following two conditions. SATO et al.: LIMITS ON THE ACCURACY OF 3-D THICKNESS MEASUREMENT IN MAGNETIC RESONANCE IMAGES 1081 (a) (a) (b) (b) (c) Fig. 5. Resected femoral head and its MR images. In subfigures (b) and (c), the horizontal and vertical axes of the images correspond, respectively, to the x axis and z axis in the left frame and the x axis and y axis in the right frame. (a) Physical appearance. Side (left) and top (right) views are shown. (b) MR images with isotropic imaging. The voxel size was : . (c) MR images with routine imaging. The voxel size was : and : . When the normal orientations were close to perpendicular to the x-y plane, i.e., , imaged cartilage appeared to be more blurred and thicker (shown by arrows). 1 = 1 5 mm (c) Fig. 4. Acrylic plate phantom and its MR images. (a) Physical appearance. (b) and (c) MR images. The horizontal and vertical axes of the images correspond to the x axis and z axis, respectively. The voxel size was : and : . As can be easily observed by naked eye, the acrylic plate with appears to be imaged slightly thicker in and than and . 1 = 1 5 mm = 1 mm =0 =0 1 = 0 625 mm = 45 =0 • Isotropic imaging—MR imaging with isotropic (cubic) [Fig. 5(b)]. Thus, voxel voxels: . anisotropy is equal to • Routine imaging—MR imaging with anisotropic (noncubic) voxels performed in clinical routine: , [Fig. 5(c)]. Thus, . voxel anisotropy is equal to ' 90 1 = 1 = 0 7 mm 1 = 0 625 mm To obtain an acceptable SNR, the imaging time with isotropic voxels was three times as long as that employed in routine imaging with anisotropic voxels. In each MR image, the matrix size was 256 256. In the obtained MR images, we observed , , and . In this case, approximately true thickness was unknown. In Section IV, we show that measured thickness with isotropic voxels can be regarded as a good approximation of true thickness under the condition and . We evaluated that thickness measured from the MR data with routine imaging by using the thickness from the MR data with isotropic imaging as a reference approximating true thickness . 3) Procedures for Thickness Determination From Real MR Images: The theory described in Section II is for continuous 1082 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 22, NO. 9, SEPTEMBER 2003 images. However, real MR images are discrete samples of continuous ones. Therefore, a thickness determination method from discrete samples is described below. Sinc interpolation: Sinc interpolation is known as the ideal interpolation to recover original continuous signal from its discrete samples, when the original signal is band-limited with respect to the sampling frequency. In order to reduce the effects of discretization, 3-D MR images were interpolated using sinc interpolation [17], [18] so that a) the signal sampling was isotropic in all three directions and b) the image matrix size was doubled [Fig. 2(b)], i.e., sinc interpolation was performed , where is the unit voxel interval for to satisfy ( , and interpolated discrete volume data set , , and are integer). It was difficult to use value smaller due to large computational cost and memory than requirement. It should be noted, however, that the volume data were inherently more blurred in the direction of than in the - plane even if the signal sampling was isotropic in all three directions, that is, the , , and directions. Detection of sheet structures: In the real MR images, detection of sheet structures was needed before thickness quantification. To facilitate the detection, sheet enhancement filtering with multiscale integration was performed according to the following equation: (38) (in which , 2, 3), and is the where denotes normal direction of the sheet at position . at combined with the second directional derivatives along Gaussian blurring of scale (SD) . In (38), was multiplied for the normalization of each scale [19], [20]. (See Appendix for .) the procedure of computing was In the experiments using the acrylic plate phantom, fixed to the known sheet normal vector irrespective of position . In the experiments using resected femoral head, where is the manually specified center position of the sphere approximating the femoral head. were thresholded, with the The filtered images threshold values being determined through operator interaction. Using connectivity analysis, the approximated segmented 3-D , were extracted. regions of the sheet structures, Thickness determination: The extracted 3-D sheet regions were thinned to a width of one voxel by nonmaximum suppression along the sheet normal directions of the sheet . A sheet thickness was assigned to enhanced images each voxel of the thinned sheet regions. For each voxel of the thinned regions, the 1-D profile of was reconstructed the second directional derivative along the sheet normal direction, where from denotes the length parameter along , and is the SD of Gaussian blurring introduced in Section II-C. (See Appendix .) Similarly, binary for detailed procedure for obtaining were reconstructed along the same profiles directional line for the binary images of the segmented sheet . The reconstruction of 1-D profiles regions and was performed at the subvoxel resolution by using a trilinear interpolation for the second directional and a nearest-neighbor interpolation for derivative the segmented sheet regions , respectively. The two sides of sheet structure edge were localized in two steps [Fig. 2(c)]: finding the initial point for the subsequent search , and searching for the zero-crossing of using . The initial point of one side of edge, , was given by if it the maximum value of that satisfied existed. Otherwise, the edge localization process terminated. The initial point of the other side of edge, , was given by the . minimum value of that satisfied Based on the initial point of the search, if , search operated toward the direction of increasing along the profile for the zero-crossing position . Otherwise search did toward , search the direction of decreasing. Similarly, if operated toward the direction of decreasing along the profile for the zero-crossing position . Otherwise search did toward the direction of increasing. The thickness, , was given by . IV. RESULTS A. Numerical Simulation The unit of dimension for the following simulation results , i.e., was assumed as described in the prewas , ) were norvious section. Thus, other parameters ( , , , and voxel anisotropy was represented as malized by . In the simulation, we mainly used and for the bar profile if not specified. These parameter values were determined so that the bar profile was symmetric and the magnitude operator in (9) did not affect the results. (The effects of the magnitude operator and asymmetrical bar profile are described in paragraph Section IV-A4 in this section.) Table I summarizes the parameter values used in the numerical simulations described below. 1) Effects of Gaussian Standard Deviation in Postprocessing: Fig. 6 shows the effects of the SD, , in Gaussian blurring. In Fig. 6(a), the relations between true thickness and , , measured thickness are shown for three values ( is equal to one, i.e., in the case of isotropic voxel. 1) when , which is the The relation is regarded as ideal when diagonal in the plots of Fig. 6(a). For each value, the relations were plotted using two values of sheet normal orientation (0 , 45 ) while was fixed to 0 . Strictly speaking, voxel shape is equal to 1 because is not perfectly isotropic even when the shape is not spherical. Thus, slight dependence on was observed. more clearly, In order to observe the deviation from . Fig. 6(b) shows the plots we defined the error as , considerable ringing of error instead of . With , error magnitude was was observed for error . With ). With , significantly large for small (around was however, ringing became small and error magnitude . gave a good comsufficiently small around promise optimizing the trade-off between reducing the ringing and improving the accuracy for small . Actually, error magniis guaranteed to satisfy for with tude , while for with and, for with . Based on this result, we used in the following experiments if not specified. SATO et al.: LIMITS ON THE ACCURACY OF 3-D THICKNESS MEASUREMENT IN MAGNETIC RESONANCE IMAGES 1083 TABLE I PARAMETER VALUES USED IN NUMERICAL SIMULATIONS (a) (b) Fig. 6. Effects of Gaussian SD, in postprocessing for thickness determination with isotropic voxel. The unit is thickness and measured thickness T . (b) Relations between true thickness and error T . 0 2) Effects of Voxel Anisotropy in MR Imaging: Fig. 7(a) and voxel shows the effects of sheet normal orientation anisotropy on measured thickness . The relations between measured thickness and sheet normal orientation for six values of true thickness (1, 2, 3, 4, 5, 6) were plotted when (1, 2, 4) were three different values of voxel anisotropy for used. The relations were regarded as ideal when any , which is the horizontal in the plots of Fig. 7(a). When , the relations were highly close to the ideal for . When and , significant deviations from the and , respectively. ideal were observed for Fig. 7(b) shows the plots of the maximum at which error is guaranteed to satisfy , , magnitude for with varied voxel anisotropy . These and plots clarify the range of where the deviation from the ideal is sufficiently small. There was no significant difference between and different values of (for ). the plots for 1 . = 0 . (a) Relations between true 3) Using Anisotropic Gaussian Blurring Based on Voxel Anisotropy: We have assumed that Gaussian blurring combined with derivative computation is isotropic as shown in (11). Another choice is to use anisotropic Gaussian blurring corresponding to voxel anisotropy, which is given by (39) and are determined so as to satisfy and, thus, because we assumed . Fig. 8(a) shows plots of measured thickness obtained using and . anisotropic Gaussian blurring when The plots using anisotropic Gaussian blurring were closer to the and any , while those using isotropic one were ideal for and . closer for 4) Effects of Bar Profile Shapes: In the above sections, the circumstances in which the magnitude operator in (9) did not affect the measurement and bar profile was symmetrical were where 1084 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 22, NO. 9, SEPTEMBER 2003 (a) 1 1 =2 1 j j 01 j j 02 2 =0 1 =1 1 =2 j j 04 =2 (b) Fig. 7. Effects of voxel anisotropy in MR imaging on measured thickness T . The unit is . = and . (a) Relations between measured thickness T and sheet normal orientation with different values. The relations are shown for three values: (left), (middle), and (right). (b) Plots of maximum at which error magnitude E is guaranteed to satisfy E < : , E < : , and E < : for while voxel anisotropy is varied (where E T ). j j = 0 1 =4 1 Fig. 8. Effects of anisotropic Gaussian blurring based on voxel anisotropy and bar profile shapes. (a) Relations between true thickness and measured thickness T with the use of anisotropic Gaussian blurring based on voxel = and = . (b) Relations between anisotropy. true thickness and measured thickness T for four different types of bar profile shapes. =2 =3 2 = (2 2)1 discussed. In order to examine the effects of the magnitude operator and asymmetrical bar profile, we further considered the and . following bar profiles. Type 0: and . Type 2: Type 1: and . Type 3: , , and . In Type 0, the profile was the same as the one used in the above presented results. The magnitude operator affected measured thickness in Types 1 and 2. The background level was in Type 1, while the bar level was (contrast was inverted) in Type 2. The bar profile was asymmetrical, , in Case 3. Fig. 8(b) shows the relations between i.e., measured thickness and sheet normal orientation for the four different types of bar profiles when voxel anisotropy was and true thickness was . The differences in measured thickness among the four types of bar profile are revealed in Fig. 8(b). While the effects of asymmetrical profiles were sufficiently small (in Type 3), those of the magnitude operator were considerable (in Types 1 and 2). Thickness was underestimated by around 0.1 in Type 1 and overestimated by around 0.1 in Type compared with Type 0. 2 for B. Validating the Numerical Simulation by In Vitro Experiments 1) Plate Phantom: Fig. 9 shows the averages and the SDs of the actually measured (in vitro) thickness from the MR data of the phantom imaged with different and and the plots of the simulated thickness representing the dependences on sheet Fig. 9. Comparison of simulated thickness and in vitro thickness determined from MR images of acrylic plate phantom. : , : , and = . For in vitro thickness, its average and SD values are indicated by symbols and error bars. (a) Dependences on sheet normal orientation . (b) Dependences on sheet normal orientation . Note that the dependence on is theoretically equivalent to the dependence on when the anisotropy is = . = (2 2)1 1 = 0 625 mm 1 = 1 5mm 1 1 =1 normal orientation and . Fig. 9(a) and (b) shows the plots of , respectively. the dependences of and with Good agreement between the simulated and the in vitro thicknesses was observed in both cases although the in vitro thicknesses was slightly greater than the simulated thickness. The biases, i.e., the difference between the simulated thickness and the average of in vitro thickness, were predominantly around 0.1 of ), and the SDs of mm or less (except for the in vitro thickness were mostly within 0.1 mm (except for of and of ). It should SATO et al.: LIMITS ON THE ACCURACY OF 3-D THICKNESS MEASUREMENT IN MAGNETIC RESONANCE IMAGES Fig. 10. Comparison of simulated thickness and in vitro thickness determined from MR images of resected femoral head. The dependence on sheet : , : , and normal orientation is shown. = . For in vitro thickness, its average and SD values are indicated by symbols and error bars, respectively. = (2 2)1 1 = 0 625 mm 1 = 1 5 mm be noted that the dependence on is theoretically equivalent to . the dependence on when the anisotropy is 2) Femoral Head Specimen: Fig. 10 shows the averages and SDs of in vitro thickness at different obtained from MR data , , and with routine imaging ( ) and the plots of the simulated thickness representing the dependence on sheet normal orientation . The averages and SDs were computed for the in vitro thickness values measured at each regardless of (because the effect of was shown to be sufficiently small in the numerical simulation). To . Let be obtain in vitro thickness, we used the reference thickness obtained from MR data with isotropic and ). The difimaging ( , ference between the reference and true thicknesses, for and was shown to be within through the numerical simulation and, thus, can be a good approximation for true thickness . We regarded the true thickness corresponding to the in vitro thickness as if the reference thickness at the same position as where was measured sat. As shown in Fig. 10, the in vitro isfied thickness was well-correlated with the simulated thickness. For of and example, underestimation around and overestimation in of were observed in the in vitro thickness as well as in the simulated thickness. The biases between in vitro and simulated thicknesses were mostly within 0.1 mm and the SDs of in vitro thickness were around 0.2 mm or less. V. DISCUSSION We have formulated a theoretical simulation method to evaluate the accuracy of thickness determination of sheet structures in 3-D MR images. Sheet structures, MR imaging processes, and postprocessing for thickness determination were modeled and simulated. One important aspect of our simulation was that these processes, which should be essentially performed in 3-D, were reduced to 1-D signal analysis in the frequency domain without loss of information. Thus, the computation of the simulation was simplified and its cost was drastically reduced. The simulation results clarified quantitative dependences of the accuracy of measured thickness on true thickness , sheet normal orientation , and bar profile shapes in sheet structures, 1085 voxel anisotropy in MR imaging, and Gaussian SD in thickness determination. For example, in order to obtain a reasonable accuracy for thickness as small as possible (e.g., satisfies for ), as error when voxel shown in Fig. 7(b), the range of should be , when , and anisotropy when using . The validity of the simulation results was confirmed by the in vitro experiments. The main findings of this work are as follows. 1) It is effective to integrate appropriate Gaussian blurring into postprocessing for thickness determination. In the simulation, was shown to be preferable for obtaining improved and stable accuracy characteristics, although without effect on noise reduc, the minimum thickness to be meation. 2) Given pixel size in sured with acceptable accuracy is approximately . Based on this finding, optimal the case of pixel size can be deduced to measure accurately the required minimum thickness. 3) The relation between voxel anisotropy and the maximum value of sheet orientation (the angle between the sheet normal and the - plane) to be measured with acceptable accuracy is shown in Fig. 7(b). Based on this result, required voxel anisotropy or slice thickness can be reasoned to measure accurately the thickness of sheet structures with desired maximum angle between the sheet normal and the - plane. 4) While there is no significant effect of asymmetrical profile of MR values along sheet normal on thickness accuracy, measured thickness is overestimated (underestimated) by around 10% in the case that the MR value of the background (sheet object) in the profile is equal or close to zero, due to the operator to obtain the magnitude from complex MR values. The effect of sheet normal orientation on measured thickness was nonmonotonic, i.e., depending on , measured thickness varied nonmonotonically with the increase of . For ex, overestimation and unample, in Fig. 7(a), when and at , derestimation were observed around , and underestimation was observed respectively, with with . In the frequency domain, around sheet structures are basically represented as the sinc function while the voxel shape determines the bandwidth to be transmitted. The sinc function has alternate negative and positive side lobes in addition to the main lobe. Depending on the bandwidth, underestimation (overestimation) occurs when negative (positive) side lobes are dominant in the transmitted frequency components. Integrating Gaussian blurring into the band-limited sinc function of sheet structures reduces the unwanted effect of the side lobes while it increases unwanted systematic overestimation in small thickness due to blurring. In the simulation, we found that a good compromise to balance this trade-off was . One important requirement in the application of Gaussian convolution is that the sampling interval of the convolution kernels needs to satisfy the Nyquest criterion. Let be the sampling interval. needs to meet in order to approximately satisfy the Nyquest criterion [19]. When and , the Nyquest criterion is not sat. Thus, isfied because it is important that the image data are interpolated without additional blurring, i.e., using sinc interpolation so that the signal 1086 sampling is isotropic in all three directions and the interval satisfies . When , , which approximately satisfies the Nyquist criterion. Previous works have addressed similar issues, which include vessel diameter measurement from MR angiography and thickness measurement from CT. These studies also evaluate the accuracy limits by using numerical simulations based on the point spread function of imaging modalities. Our work differs in the following two aspects. First, we use zero-crossings of the second directional Gaussian derivatives as the boundary definition to determined the thickness of thin sheet structures while others evaluate the diameter or thickness using the full-width at half-maximum (FWHM) [14], [21], [22]. FWHM is a useful measure for evaluating imaging systems themselves. In sophisticated image analysis systems, however, zero-crossings of the second derivatives (equivalent of the maxima of the first derivatives) combined with Gaussian blurring are needed for accurate and robust boundary localization. Our study provides comprehensive evaluations that includes computer postprocessings for thickness determination as well as imaging processes, and demonstrates the importance of parameter optimization for both imaging and postprocessing. Although zero-crossings of the second derivatives are used in [21] for the evaluation of thickness accuracy from CT images, the effects of combining Gaussian blurring are not evaluated, which is shown to have considerable effects on stabilizing the accuracy as well as on noise reduction in our study. Second, we intensively investigate the dependence of voxel anisotropy of imaging systems and the orientation of thin sheet structures. In the previous works [14], [21], [22], the tilt of vessels or thin structures relative to the plane is not evaluated. The relations among voxel anisotropy, orientation of sheet structures, and the thickness accuracy are revealed in our work. Our findings are of potential use in providing an objective criterion for optimal determination of the voxel shape and size in MR imaging, i.e., slice thickness and field of view (FOV). For example, when the distribution of orientation and thickness ranges of the concerned anatomical sheet structures are available, the slice thickness and FOV to attain acceptable accuracy with the least imaging time can be determined based on our results. Furthermore, the relations between the measured thickness and the sheet orientation [as shown in Fig. 7(a)] can be used as a calibration table to adjust the error in the measured thickness. However, the effects of noise are not addressed in the present work. It should be noted that the true thickness estimated based on the calibration table is easily affected by noise when the intervals of the plots shown in Fig. 7(a) are narrow. The results obtained in the present study are valid under the restrictions that the sheet structure has constant thickness and orientation as well as is not influenced by peripheral structures (Fig. 1). Sheet structures in in vivo MR images often violate the restrictions. For example, the brain cortex is highly curved, which results in large variations in its sheet orientation. Further, pieces of cortex are nearby each other and the femoral cartilage of the knee is nearby the tibial cartilage. In these cases, the adjacent structures influence each other. The results in this paper are not validated for the above situations. Another impor- IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 22, NO. 9, SEPTEMBER 2003 tant characteristics of cartilage assessment from MR images is the so-called “magic angle phenomenon” [23], [24], whereby its signal intensity is influenced by the angle between the static magnetic field and the local fiber orientation of the cartilage. It appears brighter at approximately 54 . Our results do not take this effect into account. In spite of the above issues, our results would give an insight into the inherent limits for the accuracy of thickness quantification in in vivo MR images. To determine the thickness from in vivo MR images, segmentation, that is, boundary detection, of the sheet structure is needed before thickness determination. There are different approaches in the boundary detection. Our results are valid for the boundary detection based on zero-crossings of the second derivatives, or equivalently maxima of the first derivatives, combined with Gaussian blurring [4], [8], [25]. When the segmentation is based on other approaches such as thresholding of original images, the results presented in this paper are not validated. However, the theoretical simulation procedures conducted in this paper can be modified so as to incorporate a different segmentation method and derive its specific accuracy characteristics. Another important aspect of this study is that the usefulness of numerical simulation in evaluating the accuracy characteristics was validated. In addition, different principles exist in the thickness definition for both sides of the detected boundaries. As long as the simple sheet model with constant thickness and orientation assumed in this paper (Fig. 1) is considered, variation in thickness definition does not affect measured thickness in general. For example, the distance along the sheet normal direction and the shortest distance between two boundaries give the same value for this simple model. However, variation in thickness definition affects measured thickness for sheet structures with large changes in thickness and orientation such as highly curved sheets. Thus, the scope of our results is basically restricted to the simple sheet model for which variation in thickness definition does not considerably affect measured thickness. VI. CONCLUSION The fundamental limits on the accuracy of thickness determination in MR images were investigated. A simulation method was established to derive the characteristics of thickness determination accuracy, considering both imaging and postprocessing parameters, especially voxel anisotropy and Gaussian blurring combined with the second derivatives. The effect of sheet structure orientation on accuracy with anisotropic (noncubic) voxels was clarified via the simulation. It was also found that postprocessing Gaussian blurring played an important role in realizing accurate and well-performed thickness determination. The simulations were validated by comparison with the actual results obtained from in vitro experiments. Future work will include the formulation of a method for automatically segmenting a sheet structure, estimating its orientation, determining its thickness, and adjusting the thickness based on the calibration table derived from simulation results. When thickness quantification is applied to in vivo MR images, the segmentation of sheet structures of interest becomes as important as the quantification itself. We are planning to develop SATO et al.: LIMITS ON THE ACCURACY OF 3-D THICKNESS MEASUREMENT IN MAGNETIC RESONANCE IMAGES a unified framework for both segmentation and quantification based on multi-scale second and first derivative analysis [25]. APPENDIX THICKNESS DETERMINATION OF SHEET STRUCTURES DISCRETE 3-D MR IMAGES IN A computationally feasible procedure for discrete MR data is described below. This procedure reproduces thickness determination for continuous images as much as possible. be 3-D MR images, in which signal sampling is Let isotropic in all three directions and its interval is . Here, , and , , and are integers. The partial second derivative combined with Gaussian blurring for the MR image with sampling interval is given by (40) where represents discrete convolution. The Hessian matrix of is given by (41) Thus, the directional second derivative along the normal direction of the sheet structure is given by (42) is satisfied, As long as the condition can be almost completely recovered from using sinc interpolation. Due to the limitation of computational cost at any and memory space, however, it is difficult to obtain continuous position . Practically, trilinear interpolation can be at any position used to obtain an approximated value of be the approxin a computationally efficient manner. Let and it is given by imated value of (43) where is the trilinear interpolation function given by (44) in which is a triangle function defined as otherwise. (45) is obtained using sinc interpolation and Similarly, discrete convolution followed by trilinear interpolation. Thickness of sheet structures can be determined by analyzing and along straight 1-D profiles of line given by (18). By substituting (18) for in and (46) and (47) 1087 are derived, respectively. Both sides of the boundaries for sheet structures can be defined as the points with the maximum and minimum values of among those satisfying the condition . The distance between the two detected given by boundary points is defined as the measured thickness, . We confirmed by experiments that the combination of sinc interpolation to reduce the sampling interval to half and discrete convolution followed by trilinear interpolation reasonably approximates the ideal continuous reconstruction for the purpose of this study. ACKNOWLEDGMENT The authors would like to thank L. Wang for making the acrylic plate phantom. REFERENCES [1] K. Jonsson, K. Buckwalter, M. Helvie, L. Niklason, and W. Martel, “Precision of hyaline cartilage thickness measurements,” Acta. Radiol., vol. 33, no. 3, pp. 234–239, 1992. [2] J. Hodler, D. Trundell, M. N. Pathria, and D. Resnick, “Width of the articular cartilage of the hip: Quantification by using fat-suppression spin-echo MR imaging in cadavers,” Amer. J. Roentgenol., vol. 159, no. 2, pp. 351–355, 1992. [3] F. Eckstein, A. Gavazzini, H. Sittek, M. Haubner, A. Losch, S. Milz, K.-H. Englmeier, E. Schulte, R. Putz, and M. Reiser, “Determination of knee joint cartilage thickness using three-dimensional magnetic resonance chondro-crassometry (3D MR-CCM),” Magn. Reson. Med., vol. 36, no. 2, pp. 256–265, 1996. [4] S. Solloway, C. E. Hutchinson, J. G. Waterton, and C. J. Taylor, “The use of active shape models for making thickness measurements of articular cartilage from MR images,” Magn. Reson. Med., vol. 37, no. 6, pp. 943–952, 1997. [5] C. A. McGibbon, D. E. Dupuy, W. E. Palmer, and D. Krebs, “Cartilage and subchondral bone thickness distribution with MR imaging,” Acad. Radiol., vol. 5, no. 1, pp. 20–25, 1998. [6] C. A. McGibbon, W. E. Palmer, and D. E. Krebs, “A general computing method for spatial cartilage thickness from co-planar MRI,” Med. Eng. Phys., vol. 20, no. 3, pp. 169–176, 1998. [7] N. Nakanishi, H. Tanaka, T. Nishii, K. Masuhara, Y. Narumi, and H. Nakamura, “MR evaluation of the articular cartilage of the femoral head during traction,” Acta. Radiol., vol. 40, no. 1, pp. 60–63, 1999. [8] Y. Sato, T. Kubota, K. Nakanishi, N. Sugano, T. Nishii, K. Ohzono, H. Nakamura, O. Ochi, and S. Tamura, “Three-dimensional reconstruction and quantification of hip joint cartilages from magnetic resonance images,” in Lecture Notes in Computer Science Berlin, Germany, 1999, vol. 1679, Proc. MICCAI’99, pp. 338–347. [9] X. Zeng, L. H. Staib, R. T. Schults, and J. S. Duncan, “Segmentation and measurement of the cortex from 3-D MR images using coupled-surfaces propagation,Segmentation and measurement of the cortex,” IEEE Trans. Med. Imag., vol. 18, pp. 927–937, Oct. 1999. [10] V. A. Magnotta, N. C. Andreasen, S. K. Schultz, G. Harris, T. Cizadlo, D. Heckel, P. Nopoulos, and M. Flaum, “Quantitative in vivo measurement of gyrification in the human brain: Changes associated with aging,” Cereb. Cortex, vol. 9, no. 2, pp. 151–160, 1999. [11] B. Fischl and A. M. Dale, “Measuring the thickness of the human cerebral cortex from magnetic resonance images,” Proc. Nat. Acad. Sci. USA, vol. 97, no. 20, pp. 11 050–11 055, 2000. [12] Y. Sato, K. Nakanishi, H. Tanaka, T. Nishii, N. Sugano, H. Nakamura, O. Ochi, and S. Tamura, “Limits to the accuracy of 3D thickness measurement in magnetic resonance images,” in Lecture Notes Computer Science Berlin, Germany, 2001, vol. 2208, Proc. MICCAI 2001, pp. 803–810. [13] D. L. Parker, Y. P. Du, and W. L. Davis, “The voxel sensitivity function in fourier transform imaging: Applications to magnetic resonance angiography,” Magn. Reson. Med., vol. 33, no. 2, pp. 156–162, 1995. [14] R. M. Hoogeveen, C. J. G. Bakker, and M. A. Viergever, “Limits to the accuracy of vessel diameter measurement in MR angiography,” JMRI—J. Magn. Reson. Imag., vol. 8, no. 6, pp. 1228–1235, 1998. [15] M. C. Steckner, D. J. Drost, and F. S. Prato, “Computing the modulation transfer function of a magnetic resonance imager,” Med. Phys., vol. 21, no. 3, pp. 483–489, 1994. 1088 [16] J. Canny, “A computational approach to edge detection,” IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-8, pp. 679–698, Nov. 1986. [17] N. M. Hylton, I. Simovsky, A. J. Li, and J. D. Hale, “Impact of section doubling on MR angiography,” Radiology, vol. 185, no. 3, pp. 899–902, 1992. [18] Y. P. Du, D. L. Parker, W. L. Davis, and G. Cao, “Reduction of partialvolume artifacts with zero-filled interpolation in three-dimensional MR angiography,” J. Magn. Reson. Imag., vol. 4, no. 5, pp. 733–741, 1995. [19] Y. Sato, S. Nakajima, N. Shiraga, H. Atsumi, S. Yoshida, T. Koller, G. Gerig, and R. Kikinis, “Three-dimensional multi-scale line filter for segmentation and visualization of curvilinear structures in medical images,” Med. Image Anal., vol. 2, no. 2, pp. 143–168, 1998. [20] Y. Sato, C.-F. Westin, A. Bhalerao, S. Nakajima, N. Shiraga, S. Tamura, and R. Kikinis, “Tissue classification based on 3-D local intensity structures for volume rendering,” IEEE Trans. Visual, Comput. Gr.aphics, vol. 6, pp. 160–180, Apr.-June 2000. IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 22, NO. 9, SEPTEMBER 2003 [21] S. Prevrhal, K. Engelke, and W. A. Kalender, “Accuracy limits for the determination of cortical width and density: The influence of object size and CT imaging parameters,” Phys. Med. Biol., vol. 44, no. 3, pp. 751–764, 1999. [22] G. Dougherty and D. Newman, “Measurement of thickness and density of thin structures by computer tomography: A simulation study,” Med. phys., vol. 26, no. 7, pp. 1341–1348, 1999. [23] S. J. Erickson, I. H. Cox, J. S. Hyde, G. F. Carrera, J. A. Strandt, and L. D. Estkowshi, “Effect of tendon orientation on MR imaging signal intensity: A manifestation of the “magic angle” phenomenon,” Radiology, vol. 181, no. 2, pp. 389–392, 1991. [24] Y. Xia, “Magic-angle effect in magnetic resonance imaging of articular cartilage: A review,” Invest. Radiol., vol. 35, no. 10, pp. 602–621, 2000. [25] Y. Sato and S. Tamura, “Detection and quantification of line and sheet structures in 3-D images,” in Lecture Notes Computer Science Berlin, Germany, 2000, vol. 1935, Proc. MICCAI 2000, pp. 154–165.
© Copyright 2026 Paperzz