Get 20M+ Full-Text Papers For Less Than $1.50/day. Subscribe now for You or Your Team.

Learn More →

Automatic determination of the arterial input function in dynamic susceptibility contrast MRI: comparison of different reproducible clustering algorithms

Automatic determination of the arterial input function in dynamic susceptibility contrast MRI:... Neuroradiology (2015) 57:535–543 DOI 10.1007/s00234-015-1493-9 FUNCTIONAL NEURORADIOLOGY Automatic determination of the arterial input function in dynamic susceptibility contrast MRI: comparison of different reproducible clustering algorithms Jiandong Yin & Jiawen Yang & Qiyong Guo Received: 5 September 2014 /Accepted: 15 January 2015 /Published online: 30 January 2015 The Author(s) 2015. This article is published with open access at Springerlink.com . . Abstract Keywords Affine propagation Arterial input function Introduction Arterial input function (AIF) plays an important Cerebral perfusion Normalized cut role in the quantification of cerebral hemodynamics. The pur- pose of this study was to select the best reproducible clustering method for AIF detection by comparing three algorithms re- Introduction ported previously in terms of detection accuracy and compu- tational complexity. In healthy individuals, alterations in neural activity are accom- Methods First, three reproducible clustering methods, nor- panied by changes in local cerebral blood flow (CBF). Thus, malized cut (Ncut), hierarchy (HIER), and fast affine propa- as Logothetis NK pointed out, CBF has been used as a marker gation (FastAP), were applied independently to simulated data of neural activity in cognitive studies [1]. Moreover, CBF which contained the true AIF. Next, a clinical verifica- plays an important part in the diagnosis and management of tion was performed where 42 subjects participated in several pathologies [2]. For example, Rau MK et al. reported dynamic susceptibility contrast MRI (DSC-MRI) scan- that dynamic susceptibility contrast (DSC)-based cerebral ning. The manual AIF and AIFs based on the different blood hemodynamics can be used to predict tumor recurrence algorithms were obtained. The performance of each al- and survival time in those patients with gliomas [3]. Recently, gorithm was evaluated based on shape parameters of the Furtner J et al. demonstrated that quantified maximum tumor estimated AIFs and the true or manual AIF. Moreover, blood flow (mTBF) values could offer a new and totally non- the execution time of each algorithm was recorded to invasive marker to prognosticate the event-free survival determine the algorithm that operated more rapidly in (EFS), independently on histopathological tumor grading, in clinical practice. patients with gliomas [4]. Therefore, cerebral perfusion is Results In terms of the detection accuracy, Ncut and linked intimately to normal and abnormal brain functions, HIER method produced similar AIF detection results, which is why numerous efforts have been made to quantify which were closer to the expected AIF and more accu- the CBF. For some urgent cases, it is also vitally important to rate than those obtained using FastAP method; in terms obtain hemodynamic information rapidly. of the computational efficiency, the Ncut method re- To minimize overexposure to ionizing radiation from some quired the shortest execution time. imaging devices, such as xenon inhalation combined with Conclusion Ncut clustering appears promising because it fa- computerized tomography and positron emission tomography, cilitates the automatic and robust determination of AIF with DSC-MRI has become a common imaging technique for ce- high accuracy and efficiency. rebral perfusion, which relies on the intravenous injection of a contrast agent and the rapid measurement of transient MRI signal changes during the passage of the bolus : : J. Yin J. Yang Q. Guo (*) through the brain [5]. Department of Radiology, Shengjing Hospital of China Medical Normally, the arterial input function (AIF) can be obtained University, No. 36, Sanhao Street, Heping District, from the mean curve of the bolus time-concentration curves Shenyang 110004, People’sRepublic of China that correspond to arterial pixels. Traditional methods for AIF e-mail: guoqy@sj-hospital.org 536 Neuroradiology (2015) 57:535–543 extraction require operators to draw a region of interest on a the three algorithms. We think that this investigation is very large artery that passes through the imaged slice, such as the important because a perfusion imaging method with improved middle cerebral artery (MCA) and internal carotid artery [6]. accuracy and speed would be beneficial for clinical use. However, the manual procedure is time-consuming and unre- peatable because it is based on operator’s experience and sub- jective judgment, which may have adverse effects on estima- Methods tion of hemodynamic parameters [7, 8]. Automatic method for AIF detection is very attractive because it is much faster, less All of the experiments were performed using an off-line personal user-dependent, and more reproducible. Thus, some automatic computer (Inter Core i3 M350 CPU processor, 2.27-GHz oper- methods have been developed. These methods include a class ating frequency, 2.0-GB RAM memory capacity, Microsoft Win- of techniques based on the cluster analysis of various multi- dows 7 operating system). A program based on MATLAB variate statistical principles, such as fuzzy c-means (FCM), k- (R2010b) was developed in our department for AIF detection. means, and hierarchy (HIER) algorithms [2, 7, 9]. Because of their high susceptibility to the randomly selected initial centers PWIdataacquisition of the clusters, both FCM and k-means reduce the calculation- recalculation reproducibility of AIFs, and further result in 1. Simulated data highly detrimental effects on disease diagnosis and tracing, so the feasibility of using these two methods for AIF detection Normally, the first passage of AIF presents the fol- is doubtful [10]. Compared with FCM and k-means, the HIER lowing shape characteristics: it is a relatively flat trend method can generate reproducible result but it is highly time- at the beginning, then, there appears a rapid shape rise consuming. Another clustering method reported by Frey and to a maximum concentration, and finally, we can see a Dueck, fast affine propagation (FastAP), was applied to dy- slower decrease after the peak. The first passage does namic contrast-enhanced MRI (DCE-MRI) data to obtain the not return to the baseline value, but it generally over- AIF, and it yielded absolutely stable results [10, 11]. However, laps with the smaller and wider peak of the second the performance of this novel method has not been evaluated recirculation contribution, mainly from the part of the for AIF detection using DSC-MRI data. In addition, another original injected bolus that was distributed to other or- more encouraging clustering method called the normalized cut gans, including the thyroid, kidneys, and lymph nodes (Ncut) algorithm is available, which uses an unbiased measure [15]. Thesimulation was performedaccordingtopub- of the disassociation between different clusters [12]. It pre- lished methods [7, 9, 16]. First, the first passage of AIF sents a good quality when minimizing the disassociation of was modeled as a gamma-variate function, as follows: different clusters, and it leads directly to maximizing the total α −ðÞ t−t =β association within the clusters [12]. ðÞ t−t  e t > t 0 0 AIFðÞ t ∝ ; ð1Þ In our opinion, the clustering methods reported previously for 0 t≤t AIF detection can be divided into two kinds, reproducible algo- rithms (Ncut, HIER, and FastAP) and irreproducible algorithms where t denotes the bolus arrival time and α and β are shape (FCM and k-means). Recently, we compared two irreproducible parameters that depend on the vasculature architecture and blood methods (k-means and FCM) [13], and we also conducted a flow, respectively. Next, a recirculation was added that comprised comparison between a reproducible method (Ncut clustering) the aforementioned AIF with a delay of τ , which was convolved and two irreproducible methods (k-means and FCM) [14]. How- with an exponential time constant τ . Based on previous studies, ever, to the best of our knowledge, there have been no compar- the parameters were set to α =3.0, β =1.5, τ =8 s,and τ =30 s d r isons between the reproducible clustering methods. Hence, it is [7, 16]. t was set to 26 s which closely approximates the arrival still a question which reproducible algorithm can obtain more time of contrast agent for our clinical perfusion data. accurate and rapid results. We think it is necessary and important According to the indicator dilution theory for intravascular to solve this problem for clinical practice. contrast agents, the time-concentration curve of the contrast To address that problem, we compared the three reproduc- agent in the tissue voxel of interest, C (t), is calculated using ible clustering methods previously used for AIF detection in the following formula: terms of the computational speed and detection accuracy. It C ðÞ t ¼ CBF  AIFðÞ t ⊗RtðÞ; ð2Þ was performed (1) on simulated data where the true AIF was known and (2) on DSC-MRI data obtained from 42 subjects where the residue function R(t) is modeled as the following where the AIF by manual drawing was extracted. To assess gamma-variate function. the performance of different methods to detect the AIF, we compared shape parameters of the estimated AIFs with the ðÞ MTT RtðÞ ¼ e ; ð3Þ reference AIFs and also compared the computational time of Neuroradiology (2015) 57:535–543 537 where MTT is the mean transit time, which equals the ratio of detailed explanation of the purpose of the study and scanning the cerebral blood volume (CBV) relative to the CBF. Ideally, procedures. the bolus is injected instantaneously and subsequently washed In total, 42 healthy volunteers participated in this study out by perfusion. (age=23–69 years; average age=49.5 years; weight=58± Next, the MRI signal intensity at time t, S(t), is obtained by 14 kg; 27 males and 15 females). The DSC-MRI data were acquired using a 3.0T whole-body MR scanner with multi- channel capabilities (MAGNETOM Verio; Siemens Medical StðÞ ¼ S  expðÞ −κ  TE  C ðÞ t ; ð4Þ 0 vox t Solution, Erlangen, Germany). A single-shot echo planar im- aging (EPI) sequence was used for perfusion imaging with the following parameters: TR=1500 ms, TE=30 ms, matrix= where S is the baseline signal intensity of the image, which 128×128, field of view (FOV)=23×23 cm, slice thickness= was 100 in this study, and κ is selected to produce a 40 % 4 mm, spacing between slices=5.2 mm, slice number=19, vox signal peak decrease from the baseline for normal gray matter acquisition type=2D, number of phase encoding steps=127, (GM), i.e., the values found in typical clinical cases [17]. The transmitting coil=body, and flip angle=90°. At the seventh time-intensity curves of the simulated MRI signal were gen- time point, a gadolinium-based contrast agent (Gadovist; erated according to the imaging form used in the clinical study Bayer Schering Pharma AG, Berlin, Germany) was adminis- mentioned below (frames, 60; echo time (TE), 0.03 s; repeti- tered using an automatic power injector at a rate of 4 ml/s, tion time (TR), 1.5 s). followed by an equal volume of saline flush at the same in- Based on Peruzzo [9], the simulated data comprised six jection speed. The temporal resolution (interval between two types of components, i.e.: adjacent frames) was 1500 ms [19]. The horizontal part of the right MCAwas covered by an imaging slice. Sixty-two whole- – Six Btrue^ arterial voxels, i.e., not affected by the partial head images were obtained (scanning duration=93 s) per sub- volume effect (PVE) ject. The magnetization state was not steady at the beginning – Sixteen Bfalse^ arterial voxels of perfusion scanning, so the first two images were discarded. – Four hundred forty voxels that simulated normal GM Therefore, 60 volumes were used for subsequent analysis. tissue – Four hundred forty voxels that simulated pathological AIF determination GM tissue – Six hundred voxels that simulated normal white matter 1. AIF calculation for simulated data (WM) tissue – Four hundred voxels contaminated by PVE First, the intensity curves of the signal were converted into contrast agent concentration curves using the following equa- These tissue states were simulated as follows [9], i.e., nor- tion [6, 7, 20, 21]: mal GM: CBV=4 ml/100 g, MTT=4±0.33 s; pathological 1 StðÞ GM: CBV=3.3 ml/100 g, MTT=10±0.7 s; and normal C ðÞ t ¼ −  ln ; ð5Þ TE  κ S vox b WM: CBV=2 ml/100 g, MTT=5.45±0.33 s. Sixteen false arterial voxels were simulated by varying t from 27 to 30 s and τ from 9 to 12 s in increments of 1.0 s. The voxels where S is the pre-contrast (baseline) signal intensity, d b contaminated by PVE were simulated using linear combina- which is obtained by averaging the pre-contrast signal tions of a true arterial signal and a signal for one of the differ- (i.e., t<26 s). ent tissues, where the weights were selected at random. Next, based on the mathematical principles outlined Finally, 400 curves were extracted randomly from the sim- by Peruzzo et al., Shi et al., and Shi et al. [9, 10, 12], ulated data and zero-mean Gaussian noise was added. The the FastAP, HIER, and Ncut methods were applied in- signal-to-noise ratio (SNR) (given by: SNR=S /σ) was set to dependently to the converted data (using Euclidean dis- 20. In general, SNR=20 is considered to be the typical noise tance). The number of voxels is the number of samples level in clinical MRI data [9, 16, 18]. for clustering. Based on previous studies, the clustering number was set to 5 [2, 9]. Normally, the time- 2. Human data concentration curves of the boluses in the arteries are characterized by a higher maximum concentration, an Ethical clearance for this study was obtained from the earlier maximum concentration, and a narrower full Ethics Committee at Shengjing Hospital, which is part of Chi- width half-maximum (FWHM), which allow the arterial na Medical University (No. 2013PS113K), and written in- curves to be distinguished from venous curves that ap- formed consent was obtained from each participant after a pear wider with later bolus arrival, and tissue curves 538 Neuroradiology (2015) 57:535–543 that are wider with a lower peak height [2, 10]. Thus, contamination by PVE, which used the ratio of the steady- to determine the AIF automatically, several parameters state integral value relative to the AUC for the first passage were calculated for the mean curve of each cluster, in- [25]. This was simplified as follows. The first passage of the cluding the maximum concentration (peak value, PV), contrast agent was fitted to a gamma variate function, and the time of maximum concentration (time to peak, TTP), area under each fitted curve was calculated and abbreviated to 1st and FWHM, as well as a measure (M)given by PV/ AUC . The beginning of the steady state was defined as the (TTP × FWHM) [2]. The cluster with the maximum M first time point that was <30 % of the maximum after the peak value was considered as arterial pixels and the AIF was of the time-concentration curve [21] and ten subsequent time obtained from this mean curve. points were integrated, which was abbreviated to SS. Finally, 1st the mean ratio of SS to AUC was calculated, and curves 2. AIF calculations using human data where the ratios fell outside the range of acceptance (mean ratio±20 %) were discarded. First, due to the misalignments of the volume images at As with the simulation study, the remaining curves different time points caused by breathing, heartbeats, and the were regarded as the input data and the Ncut, HIER, involuntary motions of the subjects, all of the volume images and FastAP methods were implemented independently. were aligned to the first pre-contrast volume based on a rigid The number of clusters was still set to 5 and the AIF transformation using SPM (available at http://www.fil.ion.ucl. cluster was again determined using the measure M=PV/ ac.uk/spm/; version SPM99) and INRIAlign 1.01 (available at (TTP × FWHM). http://www-sop.inria.fr/epidaure/Collaborations/IRMf/ Unlike the simulation study, no gold standard AIF was INRIAlign.html) [22, 23]. We did not perform the smoothing available for comparison to determine the clustering algorithm operation for any of the images. that yielded more accurate results. Thus, based on the method Second, the slice that covered the horizontal part of proposed by Shi et al. [10], the performance of each clustering the right MCA was identified manually from the first method was evaluated by comparisons with the manual meth- volume image to calculate the AIF. It has been demon- od. According to the method proposed by Mouridsen et al., six strated that the selection of this slice can result in less potential arterial voxels with the shape characteristics of an error than other slices during CBF quantification be- earlier TTP, higher PV, narrower FWHM, and quicker wash- cause of its size and its location close to brain tissue out [2, 10] were located by an experienced neural radiologist [9, 24]. As with the simulation study, the time-intensity (work experience, 34 years). The manual AIF was obtained by curves of the signals were converted into contrast agent averaging these six time-concentration curves. time-concentration curves using Eq. 5. Third, the arterial voxels comprised a minority in the se- lected slice; thus, most of the pixels represented tissue voxels Statistical analysis with small changes in signal intensity. Therefore, the area under each concentration curve (AUC) was calculated and 1. Analysis of the simulation study the P percentage of the curves with the smallest areas were AUC excluded [10]. First, the parameters of the estimated AIFs were cal- Fourth, during perfusion imaging, some fluctuating culated to evaluate the AIF detection performance of the curves were obtained because of shifts in voxels, PVEs, three different clustering methods [26]. The selected physiological pulsations, and other uncontrollable ef- AIF clusters were potentially contaminated; thus, the fects. These irregular curves would result in poor esti- PVE level was defined as the percentage of non- mates of the true AIF. Thus, a standard measurement of arterial signal in the AIF cluster, i.e., a lower PVE level curve smoothness was used, Eq. 6,and the P per- meant that the corresponding clustering algorithm could rough centage of the remaining curves with the largest integral discriminate the arterial regions well [9]. Moreover, values were excluded [2]. based on Eq. 1, we can find that the shapes of the estimated AIFs would have profound effects on subse- quent CBF quantification. Various parameters related to ∧ðÞ C ¼ ½ C ″ðÞ t dt ð6Þ the shape feature were calculated, i.e., FWHM, TTP, and PV [26]. In fact, the true AIF has the shape char- acteristics of an earlier TTP,higherPV, andnarrower As reported by Mouridsen et al., the values of P and FWHM [10]. AUC P were predefined as 0.90 and 0.25, respectively [2]. Second, the relationship between the AIF integral and CBV is rough Fifth, the novel method reported recently by Bleeker et al. a linear correlation, so the area under the AIF curve (AUC) was was applied to the remaining curves to further reduce also used to evaluate the accuracy of the estimated AIFs [9, 26]. Neuroradiology (2015) 57:535–543 539 Fig. 1 Results obtained using different clustering algorithms. a–c Arterial clusters obtained using the HIER, Ncut, and FastAP clustering methods, respectively, and d comparison of the AIFs obtained using different algorithms with the true AIF Finally, the difference between the estimated AIF and the was recorded. The statistical analysis was performed using true AIF was computed as the root mean squared error SPSS (SigmaStat, 2.03, Inc., Chicago, IL) with a t test, where (RMSE) [9]: P<0.05 was considered to indicate a significant difference. sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ½ ðÞ AIF ðÞ t −AIF ðÞ t estimate i true i i¼1 RMSE ¼ ; ð7Þ n Results where n is the scan time (90 s). Results based on the simulated data 2. Analysis of the clinical study The results are shown in Fig. 1 and Table 1. Table 1 shows that the results obtained with the Ncut meth- The true AIF curve had the shape characteristics of an od were very similar to those with the HIER method, which is earlier TTP, narrower FWHM, and higher PV [10], so the also confirmed by Fig. 1d. In contrast to the FastAP algorithm, feature parameters of the AIFs derived from different cluster- the AIFs obtained using the other two clustering methods ing methods were calculated in a similar manner to the simu- were characterized by a higher PV, earlier TTP, narrower lation study. In addition, as reported by Shi et al. [10], the error FWHM, and larger AUC and M values; thus, they were closer was also assessed using Eq. 7, where the true AIF was re- to the true AIF. This means that the Ncut and HIER methods placed by the manual AIF. Moreover, to investigate the com- are more suitable for AIF detection. The lower AIF detection putational complexity, the execution time of each algorithm performance of FastAP might have been attributable to severe Table 1 Comparison of the AIFs obtained from simulated data using different clustering methods and the true AIF AIF PVE PV TTP FWHM AUC RMSE M value HIER-based AIF 40.00 4.4030 30.01 6.3188 76.2088 0.1374 0.0232 Ncut-based AIF 45.45 4.2737 29.92 6.4563 76.4836 0.0519 0.0221 FastAP-based AIF 92.41 1.9933 31.75 27.6158 57.3931 5.5088 0.0023 True AIF 0 4.4592 29.51 6.2016 76.8669 0 0.0244 540 Neuroradiology (2015) 57:535–543 Fig. 2 AIF detection results obtained using different clustering methods. a–c Arterial clusters extracted using the HIER, Ncut, and FastAP methods, respectively, and d comparison of the manual AIF and estimated AIFs based on different algorithms PVE contamination, which is supported by Fig. 1c.According based AIF results had higher peaks than the FastAP-based to the RMSE indicator in Table 1, the Ncut-based AIF was AIF results. As noted by Enmi et al. [27], these results dem- generally more similar to the true AIF. onstrated that the HIER and Ncut methods were affected less by PVE contamination during AIF detection. This viewpoint can be appreciated intuitively based on Fig. 2c. Third, accord- Results based on the human data ing to the quantitative parameters of the shape characteristics The human AIFs were determined using the different cluster- shown in Table 2,i.e., thePV, TTP, FWHM, AUC, M value, and error, we found that the Ncut method yielded the best AIF ing methods. A randomly selected human subject was used to illustrate the differences in the AIF detection results obtained results based on the human data, which agreed with the con- clusion of the simulation study. Based on Fig. 3, we can intu- with the three clustering methods (Figs. 2 and 3, Table 2). Figure 2 shows that the results obtained using the human data itively come to the conclusion that there were only minor differences as shown. We also found that most of the arterial presented similar manner to those in the simulation study. First, the HIER- and Ncut-based AIFs had similar shape char- regions near the MCAwere detected in the tissue rather than in voxels that passed through the arteries. This finding was coun- acteristics, which were closer to the manual AIF results than the FastAP-based AIF results. Second, the HIER- and Ncut- terintuitive, but it agreed with the results reported by Bleeker Fig. 3 Detection results for the arterial regions based on different Ncut, and FastAP, respectively, where the red pixels indicate the arterial clustering methods. a Optimal slice image used for clustering analysis, regions. Compared with the FastAP algorithm, the Ncut and HIER where the horizontal part of the right MCA is indicated by the white methods could obtain more accurate arterial regions, and they increased rectangle,and b–d the AIF detection results obtained using HIER, the AIF detection performance Neuroradiology (2015) 57:535–543 541 Table 2 The parameters of AIF detection results based on different methods for the randomly selected subject AIF Shape parameters AUC M value Error Execution time (s) PV (mmol/l) TTP(s) FWHM(s) HIER-based AIF 1.4239 17.53 6.0285 18.8842 0.0135 0.03489 428.4631 Ncut-based AIF 1.5167 17.50 5.6819 19.1716 0.0153 0.03118 0.4183 FastAP-based AIF 1.187 18.80 7.3869 17.6263 0.0085 0.09036 29.9891 Manual AIF 1.6703 17.34 5.3394 19.6311 0.0180 0 107.3476 et al., where the best AIF measurements were also obtained yielding unrepeatable results [9]. Thus, they are highly disad- from voxels located in the tissue surrounding the MCA [28]. vantageous for the accurate diagnosis of patients and progres- For 42 volunteers, the pattern of differences in the AIF sion tracking. Therefore, it is necessary to develop a more shape parameters with the three methods was similar to those robust, accurate, and rapid method for automatic AIF observed in the simulation study. The detailed results of the detection. statistical analyses are shown in Table 3. Similar to the results In this study, we evaluated the AIF detection performances of the simulation study, the AIFs based on the Ncut algorithm of three previously reported clustering methods, Ncut, FastAP, had a higher PV, similar TTP, narrower FWHM, larger AUC and HIER. In contrast to traditional k-means and FCM, all of and M values, and smaller error values. There were significant the algorithms used in the present study are known to be ro- differences in the PV, FWHM, and AUC between the Ncut and bust. The results of our simulation study and clinical study FastAP methods (all P values <0.05) but not between the Ncut demonstrated that compared with the FastAP method, the Ncut and HIER methods (all P values >0.05). There was no signif- and HIER algorithms yielded AIFs more in line with the ex- icant difference in the TTP between the Ncut and HIER pected AIF, with a higher PV, earlier TTP, narrower FWHM, methods or between the Ncut and FastAP methods (all P larger AUC and M value, and smaller RMSE. The results values >0.05). In terms of the execution time, M values, and demonstrated that the FastAP method delivered low AIF de- error values, the differences between Ncut and each of the other tection performance. In addition, according to the concept of two methods reached significant level (all P values <0.05). Btime is brain,^ the time required to execute each algorithm was another important consideration in the present study. Of the two superior algorithms, the HIER method required far Discussion more time than the Ncut method and the difference was sig- nificant. For some urgent cases such as acute stroke, speed can AIF plays an important role in the quantification of cerebral save lives; thus, it is necessary to obtain the AIF results as soon perfusion using the DSC-MRI technique. The traditional man- as possible. Therefore, the Ncut method should be given pri- ual method for AIF selection is subjective, which means that ority for AIF detection in a clinical environment because of its the results lack accuracy and reproducibility between different higher detection accuracy and lower time consumption. operators, as well as between different time points. Moreover, By definition, AIF should be measured from the small ar- for some urgent cases, the AIF cannot be obtained sufficiently terioles that supply blood to the corresponding tissue, such as by rapidly using the manual procedure to make a timely deci- M3 segment of middle cerebral artery (MCA). However, be- sion [6]. Thus, the time-consuming nature of the manual meth- cause of relatively coarse spatial resolution of DSC-MRI (the od also prevents its wide application in clinical practice. The typical size is 2×2×5 mm ), a considerable PVE from the results obtained using multivariate statistical analyses based surrounding tissue would be produced, which could lead to on traditional k-means and FCM clustering algorithms are major errors in the measurement of the shape of the AIF. In highly sensitive to the initially selected cluster centers, thereby contrast, if we measure the AIF from a large artery, such as Table 3 Comparison of the AIF detection results obtained from clinical data using different clustering methods AIF Shape parameters AUC M value Error Execution time (s) PV (mmol/l) TTP (s) FWHM (s) HIER-based AIF 1.6896±0.1768 30.55±1.09 6.3165±0.8846 18.9324±1.2356 0.0142±0.0011 0.0493±0.0094 393.4381±68.3683 Ncut-based AIF 1.7395±0.1728 30.95±0.36 5.5923±0.2935 19.1081±2.6085 0.0182±0.0031 0.0397±0.0076 0.4406±0.1003 FastAP-based AIF 1.3977±0.3264 31.98±0.72 7.8908±1.3431 17.8925±3.2914 0.0076±0.0015 0.0846±0.0113 21.3792±4.1271 542 Neuroradiology (2015) 57:535–543 internal carotid artery (ICA), the measured AIF shape would References be an erroneous representation of the bolus ultimately entering the tissue. Hence, in practice, a medium size of artery, such as 1. Logothetis NK (2008) What we can do and what we cannot do with M1 segment of MCA, is commonly chosen as a compromise fMRI. Nature 453:869–878 2. Mouridsen K, Christensen S, Gyldensted L, Ostergaard L (2006) between minimizing PVE and bolus delay and dispersion [6]. Automatic selection of arterial input function using cluster analysis. It must be emphasized that there were three limitations in Magn Reson Med 55:524–531 the present study. First, only 42 subjects participated in the 3. Furtner J, Bender B, Braun C, Schittenhelm J, Skardelly M, Ernemann clinical study and this limited number of cases might have U, Bisdas S (2014) Prognostic value of blood flow measurements using arterial spin labeling in gliomas. PLoS One 9:e99616 resulted in statistical uncertainty. Thus, it will be necessary 4. Rau MK, Braun SM, Schittenhelm J, Paulsen F, Bender B, Ernemann to increase the number of subjects in a future study. Second, U, Bisdas S (2014) Prognostic value of blood flow estimated by we only calculated the global AIF. The arterial dispersion of arterial spin labeling and dynamic susceptibility contrast-enhanced the bolus between the site where the AIF was measured and MR imaging in high-grade gliomas. J Neurooncol 120:557–566 5. Roldan-Valadez E, Gonzalez-Gutierrez O, Martinez-Lopez M (2012) the true AIF could have led to an underestimation of the AIF Diagnostic performance of PWI/DWI MRI parameters in discrimi- [29]. To address this problem, the AIF of each pixel needs to nating hyperacute versus acute ischaemic stroke: finding the best be determined, i.e., the local AIF. Only a few attempts have thresholds. Clin Radiol 67:250–257 been made previously to detect the local AIF because of the 6. Calamante F (2013) Arterial input function in perfusion MRI: a com- prehensive review. Prog Nucl Magn Reson Spectrosc 74:1–32 increased risk of PVE contamination and a reduced SNR 7. Murase K, Kikuchi K, Miki H, Shimizu T, Ikezoe J (2001) [29–33]. Many different deconvolution approaches are avail- Determination of arterial input function using fuzzy clustering for able, however, which reduce the impact of tracer delay and quantification of cerebral blood flow with dynamic susceptibility dispersion, including time-insensitive block-circulant singular contrast-enhanced MR imaging. J Magn Reson Imaging 13:797–806 8. Mlynash M, Eyngorn I, Bammer R, Moseley M, Tong DC (2005) value decomposition and nonlinear stochastic regularization Automated method for generating the arterial input function on [33–38]. Third, all of the subjects who participated in this perfusion-weighted MR imaging: validation in patients with stroke. study were healthy. Thus, the clinical utility was not fully AJNR Am J Neuroradiol 26:1479–1486 validated in patients with neurological diseases, such as acute 9. Peruzzo D, Bertoldo A, Zanderigo F, Cobelli C (2011) Automatic selection of arterial input function on dynamic contrast-enhanced MR stroke, arterial stenosis, and other abnormalities. images. Comput Methods Programs Biomed 104:e148–e157 10. Shi L, Wang D, Liu W, Fang K, Wang YX, Huang W, King AD, Heng PA, Ahuja AT (2014) Automatic detection of arterial input function in dynamic contrast enhanced MRI based on affinity prop- agation clustering. J Magn Reson Imaging 39:1327–1337 Conclusion 11. Frey BJ, Dueck D (2007) Clustering by passing messages between data points. Science 315:972–976 In conclusion, we tested the suitability of three robust cluster- 12. Shi J (2000) Normalized cuts and image segmentation. IEEE Trans ing algorithms for AIF detection based on a simulation study Pattern Anal Mach Intell 22:888–905 and clinical verification. Ncut algorithm appears to be highly 13. Yin J, Sun H, Yang J, Guo Q (2014) Comparison of K-Means and fuzzy c-Means algorithm performance for automated determination promising because it facilitates automatic and robust AIF de- of the arterial input function. PLoS One 9:e85884 tection with high accuracy and low computational complexity. 14. Yin J, Sun H, Yang J, Guo Q (2014) Automated detection of the arterial input function using normalized cut clustering to determine cerebral perfusion by dynamic susceptibility contrast-magnetic reso- nance imaging. J Magn Reson Imaging Acknowledgments The authors acknowledge the Department of Radi- 15. Østergaard L (2004) Cerebral perfusion imaging by bolus tracking. ology, Shengjing Hospital, which is part of the China Medical University, Top Magn Reson Imaging 15:3–9 for providing the imaging data used in this study. We also acknowledge 16. Murase K, Shinohara M, Yamazaki Y (2001) Accuracy of the provincial government funding from the Science and Technique deconvolution analysis based on singular value decomposition for Foundation of Liaoning (No. 2011402016). quantification of cerebral blood flow using dynamic susceptibility contrast-enhanced magnetic resonance imaging. Phys Med Biol 46: Ethical Standards and patient consent We declare that this study has 3147–3159 been approved by the Ethics Committee at Shengjing Hospital, which is 17. Wu O, Østergaard L, Koroshetz WJ, Schwamm LH, O’Donnell J, part of China Medical University (No. 2013PS113K), and has therefore Schaefer PW, Rosen BR, Weisskoff RM, Sorensen AG (2003) been performed in accordance with the ethical standards laid down in the Effects of tracer arrival time on flow estimates in MR perfusion- 1964 Declaration of Helsinki and its later amendments. We declare that all weighted imaging. Magn Reson Med 50:856–864 patients gave informed consent prior to inclusion in this study. 18. Calamante F, Gadian DG, Connelly A (2003) Quantification of bolus- tracking MRI: improved characterization of the tissue residue function Conflict of interest We declare that we have no conflict of interest. using Tikhonov regularization. Magn Reson Med 50:1237–1247 19. Parker GJ, Roberts C, Macdonald A, Buonaccorsi GA, Cheung S, Open Access This article is distributed under the terms of the Creative Buckley DL, Jackson A, Watson Y, Davies K, Jayson GC (2006) Commons Attribution License which permits any use, distribution, and Experimentally-derived functional form for a population-averaged reproduction in any medium, provided the original author(s) and the high-temporal-resolution arterial input function for dynamic source are credited. contrast-enhanced MRI. Magn Reson Med 56:993–1000 Neuroradiology (2015) 57:535–543 543 20. Guzmán-de-Villoria JA, Fernández-García P, Mateos-Pérez JM, Desco 30. Grüner R, Taxt T (2006) Iterative blind deconvolution in mag- M (2012) Studying cerebral perfusion using magnetic susceptibility netic resonance brain perfusion imaging. Magn Reson Med techniques: technique and applications. Radiologia 54:208–220 55:805–815 21. Smith AM, Grandin CB, Duprez T, Mataigne F, Cosnard G (2000) 31. Willats L, Christensen S, Ma HK, Donnan GA, Connelly A, Whole brain quantitative CBF, CBV, and MTT measurements using Calamante F (2011) Validating a local arterial input function method MRI bolus tracking: implementation and application to data acquired for improved perfusion quantification in stroke. J Cereb Blood Flow from hyperacute stroke patients. J Magn Reson Imaging 2:400–410 Metab 31:2189–2198 22. Freire L, Roche A, Mangin JF (2002) What is the best similarity 32. Calamante F, Mørup M, Hansen LK (2004) Defining a local arterial measure for motion correction in fMRI time series? IEEE Trans input function for perfusion MRI using independent component anal- Med Imaging 21:470–484 ysis. Magn Reson Med 52:789–797 23. Freire L, Mangin JF (2001) Motion correction algorithms may create 33. Nordli H, Taxt T, Moen G, Grüner R (2010) Voxel-specific brain spurious brain activations in the absence of subject motion. arterial input functions from dynamic susceptibility contrast MRI Neuroimage 14:709–722 and blind deconvolution in a group of healthy males. Acta Radiol 24. Conturo TE, Akbudak E, Kotys MS, Chen ML, Chun SJ, Hsu RM, 51:334–343 Sweeney CC, Markham J (2005) Arterial input functions for dynamic 34. Wu O, Østergaard L, Weisskoff RM, Benner T, Rosen BR, Sorensen susceptibility contrast MRI: requirements and signal options. J Magn AG (2003) Tracer arrival timing-insensitive technique for estimating Reson Imaging 22:697–703 flow in MR perfusion-weighted imaging using singular value decom- 25. Bleeker EJ, van Osch MJ, Connelly A, van Buchem MA, Webb AG, position with a block-circulant deconvolution matrix. Magn Reson Calamante F (2011) New criterion to aid manual and automatic se- Med 50:164–174 lection of the arterial input function in dynamic susceptibility contrast 35. Zanderigo F, Bertoldo A, Pillonetto G, Cobelli Ast C (2009) MRI. Magn Reson Med 65:448–456 Nonlinear stochastic regularization to characterize tissue residue 26. Carroll TJ, Rowley HA, Haughton VM (2003) Automatic calculation function in bolus-tracking MRI: assessment and comparison with of the arterial input function for cerebral perfusion imaging with MR SVD, block-circulant SVD, and Tikhonov. IEEE Trans Biomed imaging. Radiology 227:593–600 Eng 56:1287–1297 27. Enmi J, Kudomi N, Hayashi T, Yamamoto A, Iguchi S, Moriguchi T, 36. Ibaraki M, Shimosegawa E, Toyoshima H, Takahashi K, Miura S, Hori Y, Koshino K, Zeniya T, Jon Shah N, Yamada N, Iida H (2012) Kanno I (2005) Tracer delay correction of cerebral blood flow with Quantitative assessment of regional cerebral blood flow by dynamic dynamic susceptibility contrast-enhanced MRI. J Cereb Blood Flow susceptibility contrast-enhanced MRI, without the need for arterial Metab 25:378–390 blood signals. Phys Med Biol 7:7873–7892 37. Mouannes-Srour JJ, Shin W, Ansari SA, Hurley MC, Vakil P, Bendok 28. Bleeker EJ, van Buchem MA, van Osch MJ (2009) Optimal location BR, Lee JL, Derdeyn CP, Carroll TJ (2012) Correction for arterial- for arterial input function measurements near the middle cerebral artery tissue delay and dispersion in absolute quantitative cerebral perfusion in first-pass perfusion MRI. J Cereb Blood Flow Metab 29:840–852 DSC MR imaging. Magn Reson Med 68:495–506 29. Knutsson L, van Westen D, Petersen ET, Bloch KM, Holtås S, 38. Mehndiratta A, MacIntosh BJ, Crane DE, Payne SJ, Chappell MA Ståhlberg F, Wirestam R (2010) Absolute quantification of cerebral (2013) A control point interpolation method for the non-parametric blood flow: correlation between dynamic susceptibility contrast MRI quantification of cerebral haemodynamics from dynamic susceptibil- and model-free arterial spin labeling. Magn Reson Imaging 28:1–7 ity contrast MRI. Neuroimage 64:560–570 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Neuroradiology Pubmed Central

Automatic determination of the arterial input function in dynamic susceptibility contrast MRI: comparison of different reproducible clustering algorithms

Neuroradiology , Volume 57 (5) – Jan 30, 2015

Loading next page...
 
/lp/pubmed-central/automatic-determination-of-the-arterial-input-function-in-dynamic-ojpyBBjqRR

References (42)

Publisher
Pubmed Central
Copyright
© The Author(s) 2015
ISSN
0028-3940
eISSN
1432-1920
DOI
10.1007/s00234-015-1493-9
Publisher site
See Article on Publisher Site

Abstract

Neuroradiology (2015) 57:535–543 DOI 10.1007/s00234-015-1493-9 FUNCTIONAL NEURORADIOLOGY Automatic determination of the arterial input function in dynamic susceptibility contrast MRI: comparison of different reproducible clustering algorithms Jiandong Yin & Jiawen Yang & Qiyong Guo Received: 5 September 2014 /Accepted: 15 January 2015 /Published online: 30 January 2015 The Author(s) 2015. This article is published with open access at Springerlink.com . . Abstract Keywords Affine propagation Arterial input function Introduction Arterial input function (AIF) plays an important Cerebral perfusion Normalized cut role in the quantification of cerebral hemodynamics. The pur- pose of this study was to select the best reproducible clustering method for AIF detection by comparing three algorithms re- Introduction ported previously in terms of detection accuracy and compu- tational complexity. In healthy individuals, alterations in neural activity are accom- Methods First, three reproducible clustering methods, nor- panied by changes in local cerebral blood flow (CBF). Thus, malized cut (Ncut), hierarchy (HIER), and fast affine propa- as Logothetis NK pointed out, CBF has been used as a marker gation (FastAP), were applied independently to simulated data of neural activity in cognitive studies [1]. Moreover, CBF which contained the true AIF. Next, a clinical verifica- plays an important part in the diagnosis and management of tion was performed where 42 subjects participated in several pathologies [2]. For example, Rau MK et al. reported dynamic susceptibility contrast MRI (DSC-MRI) scan- that dynamic susceptibility contrast (DSC)-based cerebral ning. The manual AIF and AIFs based on the different blood hemodynamics can be used to predict tumor recurrence algorithms were obtained. The performance of each al- and survival time in those patients with gliomas [3]. Recently, gorithm was evaluated based on shape parameters of the Furtner J et al. demonstrated that quantified maximum tumor estimated AIFs and the true or manual AIF. Moreover, blood flow (mTBF) values could offer a new and totally non- the execution time of each algorithm was recorded to invasive marker to prognosticate the event-free survival determine the algorithm that operated more rapidly in (EFS), independently on histopathological tumor grading, in clinical practice. patients with gliomas [4]. Therefore, cerebral perfusion is Results In terms of the detection accuracy, Ncut and linked intimately to normal and abnormal brain functions, HIER method produced similar AIF detection results, which is why numerous efforts have been made to quantify which were closer to the expected AIF and more accu- the CBF. For some urgent cases, it is also vitally important to rate than those obtained using FastAP method; in terms obtain hemodynamic information rapidly. of the computational efficiency, the Ncut method re- To minimize overexposure to ionizing radiation from some quired the shortest execution time. imaging devices, such as xenon inhalation combined with Conclusion Ncut clustering appears promising because it fa- computerized tomography and positron emission tomography, cilitates the automatic and robust determination of AIF with DSC-MRI has become a common imaging technique for ce- high accuracy and efficiency. rebral perfusion, which relies on the intravenous injection of a contrast agent and the rapid measurement of transient MRI signal changes during the passage of the bolus : : J. Yin J. Yang Q. Guo (*) through the brain [5]. Department of Radiology, Shengjing Hospital of China Medical Normally, the arterial input function (AIF) can be obtained University, No. 36, Sanhao Street, Heping District, from the mean curve of the bolus time-concentration curves Shenyang 110004, People’sRepublic of China that correspond to arterial pixels. Traditional methods for AIF e-mail: guoqy@sj-hospital.org 536 Neuroradiology (2015) 57:535–543 extraction require operators to draw a region of interest on a the three algorithms. We think that this investigation is very large artery that passes through the imaged slice, such as the important because a perfusion imaging method with improved middle cerebral artery (MCA) and internal carotid artery [6]. accuracy and speed would be beneficial for clinical use. However, the manual procedure is time-consuming and unre- peatable because it is based on operator’s experience and sub- jective judgment, which may have adverse effects on estima- Methods tion of hemodynamic parameters [7, 8]. Automatic method for AIF detection is very attractive because it is much faster, less All of the experiments were performed using an off-line personal user-dependent, and more reproducible. Thus, some automatic computer (Inter Core i3 M350 CPU processor, 2.27-GHz oper- methods have been developed. These methods include a class ating frequency, 2.0-GB RAM memory capacity, Microsoft Win- of techniques based on the cluster analysis of various multi- dows 7 operating system). A program based on MATLAB variate statistical principles, such as fuzzy c-means (FCM), k- (R2010b) was developed in our department for AIF detection. means, and hierarchy (HIER) algorithms [2, 7, 9]. Because of their high susceptibility to the randomly selected initial centers PWIdataacquisition of the clusters, both FCM and k-means reduce the calculation- recalculation reproducibility of AIFs, and further result in 1. Simulated data highly detrimental effects on disease diagnosis and tracing, so the feasibility of using these two methods for AIF detection Normally, the first passage of AIF presents the fol- is doubtful [10]. Compared with FCM and k-means, the HIER lowing shape characteristics: it is a relatively flat trend method can generate reproducible result but it is highly time- at the beginning, then, there appears a rapid shape rise consuming. Another clustering method reported by Frey and to a maximum concentration, and finally, we can see a Dueck, fast affine propagation (FastAP), was applied to dy- slower decrease after the peak. The first passage does namic contrast-enhanced MRI (DCE-MRI) data to obtain the not return to the baseline value, but it generally over- AIF, and it yielded absolutely stable results [10, 11]. However, laps with the smaller and wider peak of the second the performance of this novel method has not been evaluated recirculation contribution, mainly from the part of the for AIF detection using DSC-MRI data. In addition, another original injected bolus that was distributed to other or- more encouraging clustering method called the normalized cut gans, including the thyroid, kidneys, and lymph nodes (Ncut) algorithm is available, which uses an unbiased measure [15]. Thesimulation was performedaccordingtopub- of the disassociation between different clusters [12]. It pre- lished methods [7, 9, 16]. First, the first passage of AIF sents a good quality when minimizing the disassociation of was modeled as a gamma-variate function, as follows: different clusters, and it leads directly to maximizing the total α −ðÞ t−t =β association within the clusters [12]. ðÞ t−t  e t > t 0 0 AIFðÞ t ∝ ; ð1Þ In our opinion, the clustering methods reported previously for 0 t≤t AIF detection can be divided into two kinds, reproducible algo- rithms (Ncut, HIER, and FastAP) and irreproducible algorithms where t denotes the bolus arrival time and α and β are shape (FCM and k-means). Recently, we compared two irreproducible parameters that depend on the vasculature architecture and blood methods (k-means and FCM) [13], and we also conducted a flow, respectively. Next, a recirculation was added that comprised comparison between a reproducible method (Ncut clustering) the aforementioned AIF with a delay of τ , which was convolved and two irreproducible methods (k-means and FCM) [14]. How- with an exponential time constant τ . Based on previous studies, ever, to the best of our knowledge, there have been no compar- the parameters were set to α =3.0, β =1.5, τ =8 s,and τ =30 s d r isons between the reproducible clustering methods. Hence, it is [7, 16]. t was set to 26 s which closely approximates the arrival still a question which reproducible algorithm can obtain more time of contrast agent for our clinical perfusion data. accurate and rapid results. We think it is necessary and important According to the indicator dilution theory for intravascular to solve this problem for clinical practice. contrast agents, the time-concentration curve of the contrast To address that problem, we compared the three reproduc- agent in the tissue voxel of interest, C (t), is calculated using ible clustering methods previously used for AIF detection in the following formula: terms of the computational speed and detection accuracy. It C ðÞ t ¼ CBF  AIFðÞ t ⊗RtðÞ; ð2Þ was performed (1) on simulated data where the true AIF was known and (2) on DSC-MRI data obtained from 42 subjects where the residue function R(t) is modeled as the following where the AIF by manual drawing was extracted. To assess gamma-variate function. the performance of different methods to detect the AIF, we compared shape parameters of the estimated AIFs with the ðÞ MTT RtðÞ ¼ e ; ð3Þ reference AIFs and also compared the computational time of Neuroradiology (2015) 57:535–543 537 where MTT is the mean transit time, which equals the ratio of detailed explanation of the purpose of the study and scanning the cerebral blood volume (CBV) relative to the CBF. Ideally, procedures. the bolus is injected instantaneously and subsequently washed In total, 42 healthy volunteers participated in this study out by perfusion. (age=23–69 years; average age=49.5 years; weight=58± Next, the MRI signal intensity at time t, S(t), is obtained by 14 kg; 27 males and 15 females). The DSC-MRI data were acquired using a 3.0T whole-body MR scanner with multi- channel capabilities (MAGNETOM Verio; Siemens Medical StðÞ ¼ S  expðÞ −κ  TE  C ðÞ t ; ð4Þ 0 vox t Solution, Erlangen, Germany). A single-shot echo planar im- aging (EPI) sequence was used for perfusion imaging with the following parameters: TR=1500 ms, TE=30 ms, matrix= where S is the baseline signal intensity of the image, which 128×128, field of view (FOV)=23×23 cm, slice thickness= was 100 in this study, and κ is selected to produce a 40 % 4 mm, spacing between slices=5.2 mm, slice number=19, vox signal peak decrease from the baseline for normal gray matter acquisition type=2D, number of phase encoding steps=127, (GM), i.e., the values found in typical clinical cases [17]. The transmitting coil=body, and flip angle=90°. At the seventh time-intensity curves of the simulated MRI signal were gen- time point, a gadolinium-based contrast agent (Gadovist; erated according to the imaging form used in the clinical study Bayer Schering Pharma AG, Berlin, Germany) was adminis- mentioned below (frames, 60; echo time (TE), 0.03 s; repeti- tered using an automatic power injector at a rate of 4 ml/s, tion time (TR), 1.5 s). followed by an equal volume of saline flush at the same in- Based on Peruzzo [9], the simulated data comprised six jection speed. The temporal resolution (interval between two types of components, i.e.: adjacent frames) was 1500 ms [19]. The horizontal part of the right MCAwas covered by an imaging slice. Sixty-two whole- – Six Btrue^ arterial voxels, i.e., not affected by the partial head images were obtained (scanning duration=93 s) per sub- volume effect (PVE) ject. The magnetization state was not steady at the beginning – Sixteen Bfalse^ arterial voxels of perfusion scanning, so the first two images were discarded. – Four hundred forty voxels that simulated normal GM Therefore, 60 volumes were used for subsequent analysis. tissue – Four hundred forty voxels that simulated pathological AIF determination GM tissue – Six hundred voxels that simulated normal white matter 1. AIF calculation for simulated data (WM) tissue – Four hundred voxels contaminated by PVE First, the intensity curves of the signal were converted into contrast agent concentration curves using the following equa- These tissue states were simulated as follows [9], i.e., nor- tion [6, 7, 20, 21]: mal GM: CBV=4 ml/100 g, MTT=4±0.33 s; pathological 1 StðÞ GM: CBV=3.3 ml/100 g, MTT=10±0.7 s; and normal C ðÞ t ¼ −  ln ; ð5Þ TE  κ S vox b WM: CBV=2 ml/100 g, MTT=5.45±0.33 s. Sixteen false arterial voxels were simulated by varying t from 27 to 30 s and τ from 9 to 12 s in increments of 1.0 s. The voxels where S is the pre-contrast (baseline) signal intensity, d b contaminated by PVE were simulated using linear combina- which is obtained by averaging the pre-contrast signal tions of a true arterial signal and a signal for one of the differ- (i.e., t<26 s). ent tissues, where the weights were selected at random. Next, based on the mathematical principles outlined Finally, 400 curves were extracted randomly from the sim- by Peruzzo et al., Shi et al., and Shi et al. [9, 10, 12], ulated data and zero-mean Gaussian noise was added. The the FastAP, HIER, and Ncut methods were applied in- signal-to-noise ratio (SNR) (given by: SNR=S /σ) was set to dependently to the converted data (using Euclidean dis- 20. In general, SNR=20 is considered to be the typical noise tance). The number of voxels is the number of samples level in clinical MRI data [9, 16, 18]. for clustering. Based on previous studies, the clustering number was set to 5 [2, 9]. Normally, the time- 2. Human data concentration curves of the boluses in the arteries are characterized by a higher maximum concentration, an Ethical clearance for this study was obtained from the earlier maximum concentration, and a narrower full Ethics Committee at Shengjing Hospital, which is part of Chi- width half-maximum (FWHM), which allow the arterial na Medical University (No. 2013PS113K), and written in- curves to be distinguished from venous curves that ap- formed consent was obtained from each participant after a pear wider with later bolus arrival, and tissue curves 538 Neuroradiology (2015) 57:535–543 that are wider with a lower peak height [2, 10]. Thus, contamination by PVE, which used the ratio of the steady- to determine the AIF automatically, several parameters state integral value relative to the AUC for the first passage were calculated for the mean curve of each cluster, in- [25]. This was simplified as follows. The first passage of the cluding the maximum concentration (peak value, PV), contrast agent was fitted to a gamma variate function, and the time of maximum concentration (time to peak, TTP), area under each fitted curve was calculated and abbreviated to 1st and FWHM, as well as a measure (M)given by PV/ AUC . The beginning of the steady state was defined as the (TTP × FWHM) [2]. The cluster with the maximum M first time point that was <30 % of the maximum after the peak value was considered as arterial pixels and the AIF was of the time-concentration curve [21] and ten subsequent time obtained from this mean curve. points were integrated, which was abbreviated to SS. Finally, 1st the mean ratio of SS to AUC was calculated, and curves 2. AIF calculations using human data where the ratios fell outside the range of acceptance (mean ratio±20 %) were discarded. First, due to the misalignments of the volume images at As with the simulation study, the remaining curves different time points caused by breathing, heartbeats, and the were regarded as the input data and the Ncut, HIER, involuntary motions of the subjects, all of the volume images and FastAP methods were implemented independently. were aligned to the first pre-contrast volume based on a rigid The number of clusters was still set to 5 and the AIF transformation using SPM (available at http://www.fil.ion.ucl. cluster was again determined using the measure M=PV/ ac.uk/spm/; version SPM99) and INRIAlign 1.01 (available at (TTP × FWHM). http://www-sop.inria.fr/epidaure/Collaborations/IRMf/ Unlike the simulation study, no gold standard AIF was INRIAlign.html) [22, 23]. We did not perform the smoothing available for comparison to determine the clustering algorithm operation for any of the images. that yielded more accurate results. Thus, based on the method Second, the slice that covered the horizontal part of proposed by Shi et al. [10], the performance of each clustering the right MCA was identified manually from the first method was evaluated by comparisons with the manual meth- volume image to calculate the AIF. It has been demon- od. According to the method proposed by Mouridsen et al., six strated that the selection of this slice can result in less potential arterial voxels with the shape characteristics of an error than other slices during CBF quantification be- earlier TTP, higher PV, narrower FWHM, and quicker wash- cause of its size and its location close to brain tissue out [2, 10] were located by an experienced neural radiologist [9, 24]. As with the simulation study, the time-intensity (work experience, 34 years). The manual AIF was obtained by curves of the signals were converted into contrast agent averaging these six time-concentration curves. time-concentration curves using Eq. 5. Third, the arterial voxels comprised a minority in the se- lected slice; thus, most of the pixels represented tissue voxels Statistical analysis with small changes in signal intensity. Therefore, the area under each concentration curve (AUC) was calculated and 1. Analysis of the simulation study the P percentage of the curves with the smallest areas were AUC excluded [10]. First, the parameters of the estimated AIFs were cal- Fourth, during perfusion imaging, some fluctuating culated to evaluate the AIF detection performance of the curves were obtained because of shifts in voxels, PVEs, three different clustering methods [26]. The selected physiological pulsations, and other uncontrollable ef- AIF clusters were potentially contaminated; thus, the fects. These irregular curves would result in poor esti- PVE level was defined as the percentage of non- mates of the true AIF. Thus, a standard measurement of arterial signal in the AIF cluster, i.e., a lower PVE level curve smoothness was used, Eq. 6,and the P per- meant that the corresponding clustering algorithm could rough centage of the remaining curves with the largest integral discriminate the arterial regions well [9]. Moreover, values were excluded [2]. based on Eq. 1, we can find that the shapes of the estimated AIFs would have profound effects on subse- quent CBF quantification. Various parameters related to ∧ðÞ C ¼ ½ C ″ðÞ t dt ð6Þ the shape feature were calculated, i.e., FWHM, TTP, and PV [26]. In fact, the true AIF has the shape char- acteristics of an earlier TTP,higherPV, andnarrower As reported by Mouridsen et al., the values of P and FWHM [10]. AUC P were predefined as 0.90 and 0.25, respectively [2]. Second, the relationship between the AIF integral and CBV is rough Fifth, the novel method reported recently by Bleeker et al. a linear correlation, so the area under the AIF curve (AUC) was was applied to the remaining curves to further reduce also used to evaluate the accuracy of the estimated AIFs [9, 26]. Neuroradiology (2015) 57:535–543 539 Fig. 1 Results obtained using different clustering algorithms. a–c Arterial clusters obtained using the HIER, Ncut, and FastAP clustering methods, respectively, and d comparison of the AIFs obtained using different algorithms with the true AIF Finally, the difference between the estimated AIF and the was recorded. The statistical analysis was performed using true AIF was computed as the root mean squared error SPSS (SigmaStat, 2.03, Inc., Chicago, IL) with a t test, where (RMSE) [9]: P<0.05 was considered to indicate a significant difference. sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ½ ðÞ AIF ðÞ t −AIF ðÞ t estimate i true i i¼1 RMSE ¼ ; ð7Þ n Results where n is the scan time (90 s). Results based on the simulated data 2. Analysis of the clinical study The results are shown in Fig. 1 and Table 1. Table 1 shows that the results obtained with the Ncut meth- The true AIF curve had the shape characteristics of an od were very similar to those with the HIER method, which is earlier TTP, narrower FWHM, and higher PV [10], so the also confirmed by Fig. 1d. In contrast to the FastAP algorithm, feature parameters of the AIFs derived from different cluster- the AIFs obtained using the other two clustering methods ing methods were calculated in a similar manner to the simu- were characterized by a higher PV, earlier TTP, narrower lation study. In addition, as reported by Shi et al. [10], the error FWHM, and larger AUC and M values; thus, they were closer was also assessed using Eq. 7, where the true AIF was re- to the true AIF. This means that the Ncut and HIER methods placed by the manual AIF. Moreover, to investigate the com- are more suitable for AIF detection. The lower AIF detection putational complexity, the execution time of each algorithm performance of FastAP might have been attributable to severe Table 1 Comparison of the AIFs obtained from simulated data using different clustering methods and the true AIF AIF PVE PV TTP FWHM AUC RMSE M value HIER-based AIF 40.00 4.4030 30.01 6.3188 76.2088 0.1374 0.0232 Ncut-based AIF 45.45 4.2737 29.92 6.4563 76.4836 0.0519 0.0221 FastAP-based AIF 92.41 1.9933 31.75 27.6158 57.3931 5.5088 0.0023 True AIF 0 4.4592 29.51 6.2016 76.8669 0 0.0244 540 Neuroradiology (2015) 57:535–543 Fig. 2 AIF detection results obtained using different clustering methods. a–c Arterial clusters extracted using the HIER, Ncut, and FastAP methods, respectively, and d comparison of the manual AIF and estimated AIFs based on different algorithms PVE contamination, which is supported by Fig. 1c.According based AIF results had higher peaks than the FastAP-based to the RMSE indicator in Table 1, the Ncut-based AIF was AIF results. As noted by Enmi et al. [27], these results dem- generally more similar to the true AIF. onstrated that the HIER and Ncut methods were affected less by PVE contamination during AIF detection. This viewpoint can be appreciated intuitively based on Fig. 2c. Third, accord- Results based on the human data ing to the quantitative parameters of the shape characteristics The human AIFs were determined using the different cluster- shown in Table 2,i.e., thePV, TTP, FWHM, AUC, M value, and error, we found that the Ncut method yielded the best AIF ing methods. A randomly selected human subject was used to illustrate the differences in the AIF detection results obtained results based on the human data, which agreed with the con- clusion of the simulation study. Based on Fig. 3, we can intu- with the three clustering methods (Figs. 2 and 3, Table 2). Figure 2 shows that the results obtained using the human data itively come to the conclusion that there were only minor differences as shown. We also found that most of the arterial presented similar manner to those in the simulation study. First, the HIER- and Ncut-based AIFs had similar shape char- regions near the MCAwere detected in the tissue rather than in voxels that passed through the arteries. This finding was coun- acteristics, which were closer to the manual AIF results than the FastAP-based AIF results. Second, the HIER- and Ncut- terintuitive, but it agreed with the results reported by Bleeker Fig. 3 Detection results for the arterial regions based on different Ncut, and FastAP, respectively, where the red pixels indicate the arterial clustering methods. a Optimal slice image used for clustering analysis, regions. Compared with the FastAP algorithm, the Ncut and HIER where the horizontal part of the right MCA is indicated by the white methods could obtain more accurate arterial regions, and they increased rectangle,and b–d the AIF detection results obtained using HIER, the AIF detection performance Neuroradiology (2015) 57:535–543 541 Table 2 The parameters of AIF detection results based on different methods for the randomly selected subject AIF Shape parameters AUC M value Error Execution time (s) PV (mmol/l) TTP(s) FWHM(s) HIER-based AIF 1.4239 17.53 6.0285 18.8842 0.0135 0.03489 428.4631 Ncut-based AIF 1.5167 17.50 5.6819 19.1716 0.0153 0.03118 0.4183 FastAP-based AIF 1.187 18.80 7.3869 17.6263 0.0085 0.09036 29.9891 Manual AIF 1.6703 17.34 5.3394 19.6311 0.0180 0 107.3476 et al., where the best AIF measurements were also obtained yielding unrepeatable results [9]. Thus, they are highly disad- from voxels located in the tissue surrounding the MCA [28]. vantageous for the accurate diagnosis of patients and progres- For 42 volunteers, the pattern of differences in the AIF sion tracking. Therefore, it is necessary to develop a more shape parameters with the three methods was similar to those robust, accurate, and rapid method for automatic AIF observed in the simulation study. The detailed results of the detection. statistical analyses are shown in Table 3. Similar to the results In this study, we evaluated the AIF detection performances of the simulation study, the AIFs based on the Ncut algorithm of three previously reported clustering methods, Ncut, FastAP, had a higher PV, similar TTP, narrower FWHM, larger AUC and HIER. In contrast to traditional k-means and FCM, all of and M values, and smaller error values. There were significant the algorithms used in the present study are known to be ro- differences in the PV, FWHM, and AUC between the Ncut and bust. The results of our simulation study and clinical study FastAP methods (all P values <0.05) but not between the Ncut demonstrated that compared with the FastAP method, the Ncut and HIER methods (all P values >0.05). There was no signif- and HIER algorithms yielded AIFs more in line with the ex- icant difference in the TTP between the Ncut and HIER pected AIF, with a higher PV, earlier TTP, narrower FWHM, methods or between the Ncut and FastAP methods (all P larger AUC and M value, and smaller RMSE. The results values >0.05). In terms of the execution time, M values, and demonstrated that the FastAP method delivered low AIF de- error values, the differences between Ncut and each of the other tection performance. In addition, according to the concept of two methods reached significant level (all P values <0.05). Btime is brain,^ the time required to execute each algorithm was another important consideration in the present study. Of the two superior algorithms, the HIER method required far Discussion more time than the Ncut method and the difference was sig- nificant. For some urgent cases such as acute stroke, speed can AIF plays an important role in the quantification of cerebral save lives; thus, it is necessary to obtain the AIF results as soon perfusion using the DSC-MRI technique. The traditional man- as possible. Therefore, the Ncut method should be given pri- ual method for AIF selection is subjective, which means that ority for AIF detection in a clinical environment because of its the results lack accuracy and reproducibility between different higher detection accuracy and lower time consumption. operators, as well as between different time points. Moreover, By definition, AIF should be measured from the small ar- for some urgent cases, the AIF cannot be obtained sufficiently terioles that supply blood to the corresponding tissue, such as by rapidly using the manual procedure to make a timely deci- M3 segment of middle cerebral artery (MCA). However, be- sion [6]. Thus, the time-consuming nature of the manual meth- cause of relatively coarse spatial resolution of DSC-MRI (the od also prevents its wide application in clinical practice. The typical size is 2×2×5 mm ), a considerable PVE from the results obtained using multivariate statistical analyses based surrounding tissue would be produced, which could lead to on traditional k-means and FCM clustering algorithms are major errors in the measurement of the shape of the AIF. In highly sensitive to the initially selected cluster centers, thereby contrast, if we measure the AIF from a large artery, such as Table 3 Comparison of the AIF detection results obtained from clinical data using different clustering methods AIF Shape parameters AUC M value Error Execution time (s) PV (mmol/l) TTP (s) FWHM (s) HIER-based AIF 1.6896±0.1768 30.55±1.09 6.3165±0.8846 18.9324±1.2356 0.0142±0.0011 0.0493±0.0094 393.4381±68.3683 Ncut-based AIF 1.7395±0.1728 30.95±0.36 5.5923±0.2935 19.1081±2.6085 0.0182±0.0031 0.0397±0.0076 0.4406±0.1003 FastAP-based AIF 1.3977±0.3264 31.98±0.72 7.8908±1.3431 17.8925±3.2914 0.0076±0.0015 0.0846±0.0113 21.3792±4.1271 542 Neuroradiology (2015) 57:535–543 internal carotid artery (ICA), the measured AIF shape would References be an erroneous representation of the bolus ultimately entering the tissue. Hence, in practice, a medium size of artery, such as 1. Logothetis NK (2008) What we can do and what we cannot do with M1 segment of MCA, is commonly chosen as a compromise fMRI. Nature 453:869–878 2. Mouridsen K, Christensen S, Gyldensted L, Ostergaard L (2006) between minimizing PVE and bolus delay and dispersion [6]. Automatic selection of arterial input function using cluster analysis. It must be emphasized that there were three limitations in Magn Reson Med 55:524–531 the present study. First, only 42 subjects participated in the 3. Furtner J, Bender B, Braun C, Schittenhelm J, Skardelly M, Ernemann clinical study and this limited number of cases might have U, Bisdas S (2014) Prognostic value of blood flow measurements using arterial spin labeling in gliomas. PLoS One 9:e99616 resulted in statistical uncertainty. Thus, it will be necessary 4. Rau MK, Braun SM, Schittenhelm J, Paulsen F, Bender B, Ernemann to increase the number of subjects in a future study. Second, U, Bisdas S (2014) Prognostic value of blood flow estimated by we only calculated the global AIF. The arterial dispersion of arterial spin labeling and dynamic susceptibility contrast-enhanced the bolus between the site where the AIF was measured and MR imaging in high-grade gliomas. J Neurooncol 120:557–566 5. Roldan-Valadez E, Gonzalez-Gutierrez O, Martinez-Lopez M (2012) the true AIF could have led to an underestimation of the AIF Diagnostic performance of PWI/DWI MRI parameters in discrimi- [29]. To address this problem, the AIF of each pixel needs to nating hyperacute versus acute ischaemic stroke: finding the best be determined, i.e., the local AIF. Only a few attempts have thresholds. Clin Radiol 67:250–257 been made previously to detect the local AIF because of the 6. Calamante F (2013) Arterial input function in perfusion MRI: a com- prehensive review. Prog Nucl Magn Reson Spectrosc 74:1–32 increased risk of PVE contamination and a reduced SNR 7. Murase K, Kikuchi K, Miki H, Shimizu T, Ikezoe J (2001) [29–33]. Many different deconvolution approaches are avail- Determination of arterial input function using fuzzy clustering for able, however, which reduce the impact of tracer delay and quantification of cerebral blood flow with dynamic susceptibility dispersion, including time-insensitive block-circulant singular contrast-enhanced MR imaging. J Magn Reson Imaging 13:797–806 8. Mlynash M, Eyngorn I, Bammer R, Moseley M, Tong DC (2005) value decomposition and nonlinear stochastic regularization Automated method for generating the arterial input function on [33–38]. Third, all of the subjects who participated in this perfusion-weighted MR imaging: validation in patients with stroke. study were healthy. Thus, the clinical utility was not fully AJNR Am J Neuroradiol 26:1479–1486 validated in patients with neurological diseases, such as acute 9. Peruzzo D, Bertoldo A, Zanderigo F, Cobelli C (2011) Automatic selection of arterial input function on dynamic contrast-enhanced MR stroke, arterial stenosis, and other abnormalities. images. Comput Methods Programs Biomed 104:e148–e157 10. Shi L, Wang D, Liu W, Fang K, Wang YX, Huang W, King AD, Heng PA, Ahuja AT (2014) Automatic detection of arterial input function in dynamic contrast enhanced MRI based on affinity prop- agation clustering. J Magn Reson Imaging 39:1327–1337 Conclusion 11. Frey BJ, Dueck D (2007) Clustering by passing messages between data points. Science 315:972–976 In conclusion, we tested the suitability of three robust cluster- 12. Shi J (2000) Normalized cuts and image segmentation. IEEE Trans ing algorithms for AIF detection based on a simulation study Pattern Anal Mach Intell 22:888–905 and clinical verification. Ncut algorithm appears to be highly 13. Yin J, Sun H, Yang J, Guo Q (2014) Comparison of K-Means and fuzzy c-Means algorithm performance for automated determination promising because it facilitates automatic and robust AIF de- of the arterial input function. PLoS One 9:e85884 tection with high accuracy and low computational complexity. 14. Yin J, Sun H, Yang J, Guo Q (2014) Automated detection of the arterial input function using normalized cut clustering to determine cerebral perfusion by dynamic susceptibility contrast-magnetic reso- nance imaging. J Magn Reson Imaging Acknowledgments The authors acknowledge the Department of Radi- 15. Østergaard L (2004) Cerebral perfusion imaging by bolus tracking. ology, Shengjing Hospital, which is part of the China Medical University, Top Magn Reson Imaging 15:3–9 for providing the imaging data used in this study. We also acknowledge 16. Murase K, Shinohara M, Yamazaki Y (2001) Accuracy of the provincial government funding from the Science and Technique deconvolution analysis based on singular value decomposition for Foundation of Liaoning (No. 2011402016). quantification of cerebral blood flow using dynamic susceptibility contrast-enhanced magnetic resonance imaging. Phys Med Biol 46: Ethical Standards and patient consent We declare that this study has 3147–3159 been approved by the Ethics Committee at Shengjing Hospital, which is 17. Wu O, Østergaard L, Koroshetz WJ, Schwamm LH, O’Donnell J, part of China Medical University (No. 2013PS113K), and has therefore Schaefer PW, Rosen BR, Weisskoff RM, Sorensen AG (2003) been performed in accordance with the ethical standards laid down in the Effects of tracer arrival time on flow estimates in MR perfusion- 1964 Declaration of Helsinki and its later amendments. We declare that all weighted imaging. Magn Reson Med 50:856–864 patients gave informed consent prior to inclusion in this study. 18. Calamante F, Gadian DG, Connelly A (2003) Quantification of bolus- tracking MRI: improved characterization of the tissue residue function Conflict of interest We declare that we have no conflict of interest. using Tikhonov regularization. Magn Reson Med 50:1237–1247 19. Parker GJ, Roberts C, Macdonald A, Buonaccorsi GA, Cheung S, Open Access This article is distributed under the terms of the Creative Buckley DL, Jackson A, Watson Y, Davies K, Jayson GC (2006) Commons Attribution License which permits any use, distribution, and Experimentally-derived functional form for a population-averaged reproduction in any medium, provided the original author(s) and the high-temporal-resolution arterial input function for dynamic source are credited. contrast-enhanced MRI. Magn Reson Med 56:993–1000 Neuroradiology (2015) 57:535–543 543 20. Guzmán-de-Villoria JA, Fernández-García P, Mateos-Pérez JM, Desco 30. Grüner R, Taxt T (2006) Iterative blind deconvolution in mag- M (2012) Studying cerebral perfusion using magnetic susceptibility netic resonance brain perfusion imaging. Magn Reson Med techniques: technique and applications. Radiologia 54:208–220 55:805–815 21. Smith AM, Grandin CB, Duprez T, Mataigne F, Cosnard G (2000) 31. Willats L, Christensen S, Ma HK, Donnan GA, Connelly A, Whole brain quantitative CBF, CBV, and MTT measurements using Calamante F (2011) Validating a local arterial input function method MRI bolus tracking: implementation and application to data acquired for improved perfusion quantification in stroke. J Cereb Blood Flow from hyperacute stroke patients. J Magn Reson Imaging 2:400–410 Metab 31:2189–2198 22. Freire L, Roche A, Mangin JF (2002) What is the best similarity 32. Calamante F, Mørup M, Hansen LK (2004) Defining a local arterial measure for motion correction in fMRI time series? IEEE Trans input function for perfusion MRI using independent component anal- Med Imaging 21:470–484 ysis. Magn Reson Med 52:789–797 23. Freire L, Mangin JF (2001) Motion correction algorithms may create 33. Nordli H, Taxt T, Moen G, Grüner R (2010) Voxel-specific brain spurious brain activations in the absence of subject motion. arterial input functions from dynamic susceptibility contrast MRI Neuroimage 14:709–722 and blind deconvolution in a group of healthy males. Acta Radiol 24. Conturo TE, Akbudak E, Kotys MS, Chen ML, Chun SJ, Hsu RM, 51:334–343 Sweeney CC, Markham J (2005) Arterial input functions for dynamic 34. Wu O, Østergaard L, Weisskoff RM, Benner T, Rosen BR, Sorensen susceptibility contrast MRI: requirements and signal options. J Magn AG (2003) Tracer arrival timing-insensitive technique for estimating Reson Imaging 22:697–703 flow in MR perfusion-weighted imaging using singular value decom- 25. Bleeker EJ, van Osch MJ, Connelly A, van Buchem MA, Webb AG, position with a block-circulant deconvolution matrix. Magn Reson Calamante F (2011) New criterion to aid manual and automatic se- Med 50:164–174 lection of the arterial input function in dynamic susceptibility contrast 35. Zanderigo F, Bertoldo A, Pillonetto G, Cobelli Ast C (2009) MRI. Magn Reson Med 65:448–456 Nonlinear stochastic regularization to characterize tissue residue 26. Carroll TJ, Rowley HA, Haughton VM (2003) Automatic calculation function in bolus-tracking MRI: assessment and comparison with of the arterial input function for cerebral perfusion imaging with MR SVD, block-circulant SVD, and Tikhonov. IEEE Trans Biomed imaging. Radiology 227:593–600 Eng 56:1287–1297 27. Enmi J, Kudomi N, Hayashi T, Yamamoto A, Iguchi S, Moriguchi T, 36. Ibaraki M, Shimosegawa E, Toyoshima H, Takahashi K, Miura S, Hori Y, Koshino K, Zeniya T, Jon Shah N, Yamada N, Iida H (2012) Kanno I (2005) Tracer delay correction of cerebral blood flow with Quantitative assessment of regional cerebral blood flow by dynamic dynamic susceptibility contrast-enhanced MRI. J Cereb Blood Flow susceptibility contrast-enhanced MRI, without the need for arterial Metab 25:378–390 blood signals. Phys Med Biol 7:7873–7892 37. Mouannes-Srour JJ, Shin W, Ansari SA, Hurley MC, Vakil P, Bendok 28. Bleeker EJ, van Buchem MA, van Osch MJ (2009) Optimal location BR, Lee JL, Derdeyn CP, Carroll TJ (2012) Correction for arterial- for arterial input function measurements near the middle cerebral artery tissue delay and dispersion in absolute quantitative cerebral perfusion in first-pass perfusion MRI. J Cereb Blood Flow Metab 29:840–852 DSC MR imaging. Magn Reson Med 68:495–506 29. Knutsson L, van Westen D, Petersen ET, Bloch KM, Holtås S, 38. Mehndiratta A, MacIntosh BJ, Crane DE, Payne SJ, Chappell MA Ståhlberg F, Wirestam R (2010) Absolute quantification of cerebral (2013) A control point interpolation method for the non-parametric blood flow: correlation between dynamic susceptibility contrast MRI quantification of cerebral haemodynamics from dynamic susceptibil- and model-free arterial spin labeling. Magn Reson Imaging 28:1–7 ity contrast MRI. Neuroimage 64:560–570

Journal

NeuroradiologyPubmed Central

Published: Jan 30, 2015

There are no references for this article.