Eur Radiol DOI 10.1007/s00330-017-4773-4
NEURO
Diagnosis of small posterior fossa stroke on brain CT: effect of iterative reconstruction designed for brain CT on detection performance Taihei Inoue 1 & Takeshi Nakaura 2 & Morikatsu Yoshida 1 & Koichi Yokoyama 1 & Kenichiro Hirata 2 & Masafumi Kidoh 2 & Seitaro Oda 2 & Daisuke Utsunomiya 2 & Kazunori Harada 3 & Yasuyuki Yamashita 2
Received: 8 August 2016 / Revised: 30 January 2017 / Accepted: 9 February 2017 # European Society of Radiology 2017
Abstract Objectives In this study, we aimed to determine whether iterative model reconstruction designed for brain CT (IMRneuro) would improve the accuracy of posterior fossa stroke diagnosis on brain CT. Methods We enrolled 37 patients with ischaemic stroke in the posterior fossa and 37 patients without stroke (controls). Using axial images reconstructed using filtered backprojection (FBP) and IMR-neuro, we compared the CT numbers in infarcted areas, image noise in the pons, and contrastto-noise ratios (CNRs) of infarcted and non-infarcted areas on scans subjected to IMR-neuro and FBP. To analyse the performance of hypo-attenuation detection, we used receiveroperating characteristic (ROC) curve techniques. Results The image noise was significantly lower (2.2 ± 0.5 vs. 5.1 ± 0.9 Hounsfield units, p < 0.01) and the difference in CNR between the infarcted and non-infarcted areas was significantly higher with IMR-neuro than with FBP (2.2 ± 1.7 vs. 4.0 ± 3.6, p < 0.01). Furthermore, the average area under the ROC curve was significantly higher with IMR-neuro (0.90 vs. 0.86 for FBP, p = 0.04). Conclusion IMR-neuro yielded better image quality and improved hypo-attenuation detection in patients with ischaemic stroke.
* Taihei Inoue
[email protected]
1
Department of Radiology, Amakusa Medical Center, 854-1 Jikiba, Amakusa, Kumamoto, Japan
2
Department of Diagnostic Radiology, Graduate School of Life Sciences, Kumamoto University, Kumamoto, Japan
3
Department of Surgery, Amakusa Medical Center, Kumamoto, Japan
Key points • Iterative model reconstruction of brain CT data can facilitate the diagnosis of ischaemic stroke. • IMR improved the detectability of low-contrast lesions in the posterior fossa. • IMR-neuro yielded better image quality and improved observer performance. Keywords Computed tomography . Image processing . Stroke . Radiation dosage . Posterior fossa
Introduction Brain computed tomography (CT) is convenient and feasible for investigating both the acute and chronic phases of ischaemic stroke [1]. However, it is difficult to differentiate small acute ischaemic strokes from older infarctions in the posterior fossa. CT scanners generally use calibration correction and iterative beam-hardening correction software to minimize the beam-hardening effects of cranial bone, which may result in overcorrection or undercorrection depending on the anatomical complexity [2]. Furthermore, it is more difficult to detect ischaemic lesions in the posterior fossa relative to the cerebrum, given the smaller volume of the former and increased cranial bone thickness [3]. Consequently, image noise and beam-hardening artifacts should be reduced to improve lesion visualization in the posterior fossa. Although high-tube-current CT can reduce quantum image noise [4], it also increases radiation exposure. Accordingly, a model-based iterative reconstruction (MBIR) algorithm that features iterative forward- and back-projection and more sophisticated modeling was introduced to improve image quality [5–9]. However, early MBIR techniques included only one
Eur Radiol
fixed reconstruction parameter and were not optimized for brain CT. A newly available MBIR technique, iterative model reconstruction (IMR, Philips Healthcare, Cleveland, OH, USA; knowledge-based iterative model reconstruction) which includes head-specific techniques (IMR-neuro), adjusts the iterative reconstruction parameters of space and contrast resolution according to user input and body area, using penalty factors via the cost function. Like conventional reconstruction kernels used in FBP, the IMR parameters are tuned to reduce beam-hardening artefacts and improve visibility of the brain parenchyma at the gray–white matter interface [10]. To our knowledge, no clinical studies of the diagnostic performance of IMR-neuro for the detection of hypodense lesions of the brainstem and cerebellum have been published. Accordingly, in this study, we aimed to evaluate the usefulness of MBIR specialized for brain CT for the detection of ischaemic lesions in these brain regions.
Materials and methods This retrospective study was approved by our institutional review board, and the requirement for informed patient consent was waived. Patients We reviewed the records of 146 patients who underwent brain CT studies with a 256-slice multidetector scanner (MDCT) at Amakusa Medical Center between April 2014 and May 2016. Of them, 115 patients had undergone magnetic resonance imaging (MRI); of these patients, 37 were confirmed to have experienced brain infarction in the posterior fossa and were classified as the infarction group. Other patients were classified as the non-infarction group (Fig. 1). We recorded the times from symptom onset of patients in the infarction group Fig. 1 Flow diagram of the patient selection process
and evaluated infarction areas using the pc-ASPECTS scoring system [11]. To reduce potential confounding effects, we selected 37 control patients using propensity score-based matching analysis. In these patients, brain CT imaging was primarily performed to address patient complaints of headache and dizziness, and MRI indicated no obvious changes to the brain parenchyma. Patient age and sex were included in the propensity score calculations. Continuous propensity scores ranging from 0 to 1 were calculated using a logistic regression analysis, and one-to-one matching without replacement was performed by 0.05 caliper matching of the estimated propensity scores. We used the JMP statistical software package (version 12; SAS, Cary, NC, USA) for the propensity score-based matching analysis. CT scanning and reconstruction All patients were scanned on a 256-slice MDCT device (Brilliance iCT, Philips Healthcare, Cleveland, OH, USA) with the following parameters: 120 kVp; gantry rotation, 0.4 s; helical pitch, 0.39; mA tube current, 293 mA; and CT volume dose index (CTDIvol), 40.8 mGy (Table 1). Image reconstruction was performed in a 25-cm display field of view (FOV), with a slice thickness of 5 mm. All images were reconstructed using a standard (FBP) algorithm with a standard brain kernel (UB) or the IMR-neuro technique (level 1, the nominal setting suggested by the vendor for brain imaging). Quantitative image analysis A board-certified radiologist with 4 years of experience reading brain CT scans subjected images obtained using the two reconstruction techniques to a quantitative image analysis. Areas corresponding to brainstem or cerebellar
Eur Radiol Table 1 Scan parameters of brain CT
Scan parameter Beam collimation (mm) Slice thickness (mm) Tube voltage (kVp) Tube current (mA) Rotation time (sec) CTDIvol (mGy)
128 × 0.625 5 120 293 0.4 40.8
CTDI CT dose index
sequences. The mean attenuations of corresponding areas on CT images were also recorded. If possible, a 25 mm2 region of interest (ROI) was placed in the centre of the infarcted area; otherwise, the widest ROI possible was used. Another circular ROI (~50 mm2) was placed in the non-infarcted area. To evaluate image noise, we measured the standard deviation (SD) of attenuation in a normal area of the brainstem (ROI: ~80 mm2). Image noise was defined as the SD of the ROI in the brainstem. The image contrast and contrast-to-noise ratio (CNR) for each reconstruction technique were calculated as follows:
infarctions were identified on diffusion-weighted (DW) or fluid-attenuated inversion recovery (FLAIR) MRI Contrast ¼ C T number of n o n − infarcted R O I – C T number of infarcted R O I CNR ¼ ðCT number of non−infarcted ROI – CT number of infarcted ROIÞ=ðimage noiseÞ
Review protocol Six radiologists with 7–19 years of clinical neuroimaging review experience and >3 months experience with IMR image review participated in the observer performance study. Prior to the review, we prepared and stored anonymised Digital Imaging and Communications in Medicine files that contained CT images reconstructed using FBP or IMR-neuro for each study subject. The IMR-neuro technique has increased computational requirements and reduced recon speeds, compared with the FBP technique. However, graphics processing units allowed us to shorten the reconstruction time to a clinically acceptable duration (<3 min for brain CT). An author (T.I.) who did not participate in the image review selected one image from each DICOM dataset. In the infarction group (n = 37), these images displayed the acute or chronic stroke in the brainstem or cerebellum, whereas in the control group (n = 37) the images showed the mid-brainstem and the cerebellum. After the dataset (i.e. FBP and IMR-Neuro reconstructions for control and infarction groups) order
was randomized, each reviewer evaluated all 72 datasets. The receiver-operating characteristic (ROC) curve analysis was performed at a window level of 40 Hounsfield units (HU) and width of 80 HU (standard brain window setting) on a PACS viewer (Synapse; FUJIFILM Medical Co. Ltd., Tokyo, Japan). The observers were blinded to the acquisition parameters, clinical symptoms, MRI findings and patients' ages, sexes and follow-up clinical findings. They indicated their level of confidence regarding the presence or absence of parenchymal hypo-attenuation by placing a mark on a continuous rating scale [12]. The reading time was not limited. Statistical analysis All numeric values are reported as means ± SDs. The CT number, contrast, image noise and CNR on scans subjected to FBP and IMR-neuro were compared using Student’s t-test. The effect of each reconstruction technique on the detection of parenchymal hypo-attenuation was assessed using an ROC curve analysis, and the
Table 2 Quantitative image analysis Mean CT number – non-infarcted area (HU) Mean CT number – infarcted area (HU) Contrast (HU) Image noise (HU) CNR
FBP
IMR
P value FBP vs. IMR
29.0 ± 5.1 19.0 ± 5.5 10.5 ± 7.2
28.5 ± 4.6 19.7 ± 4.9 8.9 ± 6.7
>0.66 >0.59 0.31
5.1 ± 0.9 2.2 ± 1.7
2.2 ± 0.5 4.0 ± 3.6
<0.01 <0.01
Data are mean ± standard deviation HU Hounsfield unit, FBP filtered back projection, IMR iterative model reconstruction, CNR contrast to noise ratio
Eur Radiol Fig. 2 Box plots of the quantitative analysis of (a) the image contrast, (b) the image noise, and (c) the contrast-tonoise ratio (CNR) on scans obtained using both reconstruction methods. With iterative model reconstructionneuro (IMR) the CNR was higher than with filtered back projection (FBP) (P < 0.03)
ROCKIT 0.9B software program (C. E. Metz, University of Chicago, Chicago, IL, USA) was used to generate binomial ROC curves from continuous rating data [12]. We tested the statistical significance of differences in the areas under the curves (AUCs) obtained with the two reconstruction techniques via the jack-knife method for generalizing the population of readers and patients (including reader- and case-sample variations). A P value <0.05 was considered significant. We also used an analysis of variance approach [13] enabled by the DBM MRMC 2.2 program (C. E. Metz, University of Chicago, Chicago, IL, USA). Statistical analyses were performed using the free statistical software package R (version 3.2.1; The R Project for Statistical Computing; http://www.r-project.org/).
Conventional Binormal ROC Curves 1
Results Patient demographics Five, 18 and 14 of 37 patients underwent brain CT imaging within 4.5 h, from 4.5 h to 1 week, and after 1 week from symptom onset, respectively. The pc-ASPECTS scoring system was used to evaluate the degree of ischaemia in all patients [11]. Of the 37 patients, 12, 16, three and four received scores of 9, 8, 7 and 6, respectively. Quantitative image analysis As shown in Table 2, no significant differences related to reconstruction type were observed in the CT number and contrast between ROIs placed in infarcted and non-infarcted areas (p > 0.05). However, IMR-neuro yielded significantly lower image noise (p < 0.01) and a significantly higher CNR between infarcted and non-infarcted areas (p < 0.01) (Fig. 2). ROC analysis
0.8
Figure 3 shows the averaged alternative free-response ROC curves used to detect parenchymal hypo-attenuation. The Az values for the six observers are listed in Table 3. The mean Az values recorded for IMR-neuro images were significantly
TPF
0.6
FBP Az=0.86
0.4
IMR Az=0.90
Table 3 Az values for performance in detecting parenchymal hypoattenuation
0.2
0 0
0.2
0.4
0.6
0.8
1
FPF
Fig. 3 Receiver operating characteristic curves showing the diagnostic performance of filtered back projection (FBP) and iterative model reconstruction-neuro (IMR-neuro) in patients with hypo-dense areas in the posterior fossa. The diagnostic performance of IMR-neuro was significantly higher than of FBP
Az value FBP
IMR
Reader 1 Reader 2 Reader 3
0.89 0.90 0.78
0.92 0.90 0.87
Reader 4 Reader 5 Reader 6
0.88 0.85 0.84
0.91 0.91 0.89
FBP filtered back projection, IMR iterative model reconstruction Figure Legends
Eur Radiol
Fig. 4 A 93-year-old woman with acute infarction in the left pons. In this case, pc-ASPECTS value was 8. The images were acquired on the day after symptom onset. With iterative model reconstruction-neuro (IMRneuro) (b) the image noise was significantly lower than with filtered back
projection (FBP) (a). On IMR-neuro images the subtle parenchymal hypo-attenuation of the infarcted area was clearly depicted (arrow). On the follow-up diffusion-weighted (DW) MRI scan (c) the ischaemic stroke focus is clearly demonstrated
higher than those recorded for FBP images (0.90 ± 0.10 vs. 0.86 ± 0.13, p = 0.04). Representative cases are shown in Figs. 4 and 5.
In this study of patients with ischaemic stroke in the posterior fossa, we demonstrated that the use of IMR-neuro, rather than FBP, led to a significant decrease in image noise and significant increase in the CNR of the infarcted area. IMR-neuro also improved observer performance regarding the detection of parenchymal hypo-attenuation in the posterior fossa. The American Heart Association currently recommends non-enhanced CT as the initial modality for investigations of ischaemic stroke [14] because it is more rapid and feasible than MRI in most institutions; however, this modality is less accurate than DW MRI [15, 16]. Following ischaemic stroke, the decrease in CT density in the infarcted brain parenchyma is relatively small, and the inherent image noise on brain CT scans would hinder the delineation of such small attenuation changes [12, 13]. Furthermore, severe beam-hardening effects in the posterior fossa can easily mask small lesions. However, improved CT diagnostic abilities are needed to address the potentially fatal nature of small brainstem lesions.
Although the MBIR method is suitable for most imaging modalities, it has not been widely applied to CT studies of the posterior fossa, as the reconstruction of CT images of this area requires an accurate assessment of the skull bone beamhardening effect to avoid artefacts. In FBP-based methods, bone-related beam-hardening is generally corrected with forward- and back-projection during post-processing [17]. However, this process increases the difficulty of developing a full-IR method for brain studies. Here, we have demonstrated that MBIR specialized for brain CT improved both the image quality and the ability to detect low-contrast lesions in the posterior fossa. However, other researchers reported that IR failed to improve the detectability of low-contrast objects at low radiation doses [18, 19]. We propose two possibilities for this discrepancy. First, our study used a higher radiation dose, compared to those used in earlier phantom studies that reported no marked differences in low-contrast detectability on IR- and FBP images [19, 20]. The spatial resolution of MBIR depends on both the radiation dose and image contrast level, and can be degraded at low radiation dose settings, particularly when imaging low-contrast objects [21, 22]. Second, most studies involving MBIR used a single fixed reconstruction parameter; in contrast, we used a MBIR protocol specialized for brain CT. We note that no publications have compared the image quality of scans acquired via MBIR with
Fig. 5 An 85-year-old man presenting with headache. In this case, pcASPECTS value was 8. Note dilation of the perivascular space on the right side of the pons. With iterative model reconstruction-neuro (IMRneuro) (b) the image noise was significantly lower than with filtered back
projection (FBP) (a) and the subtle parenchymal hypo-attenuation of the infarcted area is clearly depicted (arrow). On the T2-weighted (T2WI) MRI scan acquired before CT for another purpose (c) this area was hyperintense and thought to be cerebrospinal fluid
Discussion
Eur Radiol
a single fixed parameter versus MBIR specialized for brain CT. However, Nakaura et al. [10], in an evaluation of the image qualities yielded by FBP, hybrid-IR, and MBIR specialized for brain CT, suggested that the latter provided better image quality, compared to hybrid-IR. We suggest that optimization might improve the detectability of subtle hypodense areas in the posterior fossa. Finally, a study of acute ischaemic stroke in the posterior fossa might be more interesting. However, we did not have a sufficient population of acute ischaemic stroke patients for a comparative study. Further studies are therefore needed to confirm our findings regarding acute ischaemic patients. Our study has some limitations. First, the small population size limits the generalisability of our findings. Second, we performed a conventional quantitative analysis of image noise and CNR and observed that the low-contrast resolution of FBP reconstruction improved as the noise decreased because the spatial resolution remained the same, regardless of the object contrast. In contrast, when using the IR technique, the spatial resolution changes as a function of the object. Therefore, the image noise and CNR might not be appropriate measures of low-contrast resolution when using IR techniques. However, there is currently no widely accepted method for evaluating the low-contrast resolution of clinical CT images reconstructed via IR techniques. Finally, although the readers were blinded to the reconstruction methods, experienced readers could easily identify which technique had been used to reconstruct each CT image. Therefore, we could not completely blind the reconstruction methods. In conclusion, when compared with FBP, MBIR specialized for brain CT yielded improved image quality and improved reader performance for the detection of parenchymal hypo-attenuation in the posterior fossa. Acknowledgements The scientific guarantor of this publication is Yasuyuki Yamashita. The authors of this manuscript declare no relationships with any companies whose products or services may be related to the subject matter of the article. The authors state that this work has not received any funding. No complex statistical methods were necessary for this paper. Written informed consent was waived by the Institutional Review Board. Institutional Review Board approval was obtained. Methodology: retrospective, diagnostic or prognostic study, performed at one institution.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14. 15.
16.
17. 18.
19.
References 20. 1.
2.
3.
Bernick C, Kuller L, Dulberg C et al (2001) Silent MRI infarcts and the risk of future stroke: the cardiovascular health study. Neurology 57:1222–1229 Joseph PM, Ruth C (1997) A method for simultaneous correction of spectrum hardening artifacts in CT images containing both bone and iodine. Med Phys 24:1629–1634 Rozeik C, Kotterer O, Preiss J, Schutz M, Dingler W, Deininger HK (1991) Cranial CT artifacts and gantry angulation. J Comput Assist Tomogr 15:381–386
21.
22.
Ogawa A, Mori E, Minematsu K et al (2007) Randomized trial of intraarterial infusion of urokinase within 6 hours of middle cerebral artery stroke: the middle cerebral artery embolism local fibrinolytic intervention trial (MELT) Japan. Stroke 38:2633–2639 Block KT, Uecker M, Frahm J (2009) Model-based iterative reconstruction for radial fast spin-echo MRI. IEEE Trans Med Imaging 28:1759–1769 Deak Z, Grimm JM, Treitl M et al (2013) Filtered back projection, adaptive statistical iterative reconstruction, and a model-based iterative reconstruction in abdominal CT: an experimental clinical study. Radiology 266:197–206 Pickhardt PJ, Lubner MG, Kim DH et al (2012) Abdominal CT with model-based iterative reconstruction (MBIR): initial results of a prospective trial comparing ultralow-dose with standard-dose imaging. AJR Am J Roentgenol 199:1266–1274 Katsura M, Matsuda I, Akahane M et al (2012) Model-based iterative reconstruction technique for radiation dose reduction in chest CT: comparison with the adaptive statistical iterative reconstruction technique. Eur Radiol 22:1613–1623 Chang W, Lee JM, Lee K et al (2013) Assessment of a modelbased, iterative reconstruction algorithm (MBIR) regarding image quality and dose reduction in liver computed tomography. Invest Radiol 48:598–606 Nakaura T, Iyama Y, Kidoh M et al (2016) Comparison of iterative model, hybrid iterative, and filtered back projection reconstruction techniques in low-dose brain CT: impact of thin-slice imaging. Neuroradiology 58:245–251 Puetz V, Sylaja PN, Coutts SB et al (2008) Extent of hypoattenuation on CT angiography source images predicts functional outcome in patients with basilar artery occlusion. Stroke 39:2485–2490 Koc G, Courtier JL, Phelps A, Marcovici PA, MacKenzie JD (2014) Computed tomography depiction of small pediatric vessels with model-based iterative reconstruction. Pediatr Radiol 44:787–794 Dzialowski I, Weber J, Doerfler A, Forsting M, von Kummer R (2004) Brain tissue water uptake after middle cerebral artery occlusion assessed with CT. J Neuroimaging 14:42–48 Brooks RA, Di Chiro G (1976) Statistical limitations in x-ray reconstructive tomography. Med Phys 3:237–240 Barber PA, Demchuk AM, Zhang J, Buchan AM (2000) Validity and reliability of a quantitative computed tomography score in predicting outcome of hyperacute stroke before thrombolytic therapy. ASPECTS Study Group. Alberta Stroke Programme Early CT Score Lancet 355:1670–1674 Haubenreisser H, Fink C, Nance JW Jr et al (2014) Feasibility of slice width reduction for spiral cranial computed tomography using iterative image reconstruction. Eur J Radiol 83:964–969 Kijewski PK, Bjarngard BE (1978) Correction for beam hardening in computed tomography. Med Phys 5:209–214 Kidoh M, Nakaura T, Awai K et al (2013) Novel connecting tube for saline chaser in contrast-enhanced CT: the effect of spiral flow of saline on contrast enhancement. Eur Radiol 23:3213–3218 Nakaura T, Kidoh M, Sakaino N et al (2013) Low radiation dose protocol in cardiac CT with 100 kVp: usefulness of display preset optimization. Int J Cardiovasc Imaging 29:1381–1389 Kidoh M, Nakaura T, Nakamura S et al (2014) Reduction of dental metallic artifacts in CT: value of a newly developed algorithm for metal artifact reduction (O-MAR). Clin Radiol 69:e11–e16 Kidoh M, Nakaura T, Nakamura S et al (2014) Low-contrast-dose protocol in cardiac CT: 20% contrast dose reduction using 100 kVp and high-tube-current-time setting in 256-slice CT. Acta Radiol 55: 545–553 Nakaura T, Nagayoshi Y, Awai K et al (2014) Myocardial bridging is associated with coronary atherosclerosis in the segment proximal to the site of bridging. J Cardiol 63:134–139