Visual Cerebral Microbleed Detection on 7T MR Imaging: Reliability and Effects of Image Processing

SUMMARY: MR imaging at 7T has a high sensitivity for cerebral microbleed detection. We identified mIP processing conditions with an optimal balance between the number of visually detected microbleeds and the number of sections on 7T MR imaging. Even with optimal mIP processing, the limited size of some of the microbleeds and the susceptibility effects of other adjacent structures were a challenge for visual detection, which led to a modest inter-rater agreement, mainly due to missed microbleeds. Automated lesion-detection techniques may be required to optimally benefit from the increased spatial resolution offered by 7T MR imaging.

T here is a growing interest in cerebral microbleeds on MR imaging in the context of aging, cerebrovascular disease, and dementia. Recent studies indicate that microbleeds may be markers of specific underlying microvascular pathologies. Furthermore, the presence of microbleeds may convey prognostic information in patients with stroke or dementia. Improvement in the detection of microbleeds is important to further our understanding of this novel marker of cerebral small-vessel disease.
The sensitivity to detect microbleeds increases with higher MR imaging field strengths. 1 On 7T dual-echo T2*-weighted MR imaging, for example, up to 3 times as many microbleeds can be detected than on 1.5T MR images. 2 However, susceptibility effects of other structures also complicate microbleed detection at higher field strengths. 1 We sought to determine mIP processing conditions with an optimal balance between accuracy and rating time and to determine intra-and interrater agreement for these processing conditions on 7T MR imaging.

Subjects
Subjects were selected from the SMART 7T MR imaging study. 2 The SMART study is a prospective cohort study that assesses brain changes in patients with manifest arterial disease. For the comparison of the different mIP conditions, we first selected scans from 3 subjects (mean age, 72 Ϯ 8 years) who each had at least 10 microbleeds on their 7T MR images, based on previous evaluations by raters who were not involved in the present study.
Subsequently, scans from 10 additional subjects (8 with and 2 without microbleeds; mean age, 67 Ϯ 8 years; the raters were not aware of the number of subjects with and without microbleeds) were selected to study intra-and inter-rater agreement. This study was approved by the local ethics committee of the University Medical Center Utrecht, and all subjects signed an informed consent form.

MR Imaging Scanning Protocol
Scans were acquired on a 7T MR system (Philips Healthcare, Cleveland, Ohio) with a volume-transmit and 16-channel receive head coil (Nova Medical, Wilmington, Massachusetts). Dualecho 3D T2*-weighted images were obtained (TR/first TE/second TE ϭ 20/2. Eight of the 10 subjects also had a 1.5T MR imaging scan (Gyroscan ACS-NT; Philips Healthcare, Best, the Netherlands) with 3D T2*-weighted images (TR/TE ϭ 25/35 ms) with an acquired voxel size of 0.8 ϫ 0.8 ϫ 0.8 mm 3 (for details see Conijn et al 2 ) and T2-weighted images (TR/TE ϭ 2200/100 ms) with an acquired voxel size of 0.90 ϫ 1.03 ϫ 4.00 mm 3 on the same day as the 7T scan.
Because the 2/0 and 4/Ϫ2 mIP condition were found to have an optimal tradeoff between the number of detected microbleeds and the number of sections, these conditions were selected to assess intra-and inter-rater agreement and the effect of overlap.

Microbleed Rating
For the 7T scans, microbleeds were rated with concomitant onscreen presentation of the first-and second-echo mIP processed images. Ratings were performed by the MARS. 3 Microbleeds were defined as black, round, or ovoid lesions on the first-or secondecho image and distinguished into possible or definite microbleeds. 3 Mimics, like vessels and symmetric calcifications in the basal ganglia and choroid plexus, were excluded. 3 FLAIR or T2weighted images were used to detect infarcts. Microbleeds were not rated within infarcted brain tissue. We made 2 modifications to the MARS for application on the 7T scans. First, we used no lower size limit for the lesions (2 mm is the lower size limit of MARS). Second, lesions visible on the first-echo image had to be larger on the second-echo image (so-called blooming) to be rated as microbleeds. 2 Before the formal rating, the raters practiced the rating method together on 7T scans of individuals from the SMART study who were not included in the present study. Levels of experience with brain MR imaging rating varied across raters. Nevertheless, on the 1.5T scans intra-and inter-rater agreement was excellent (

Ground Truth Ratings
To assess the effect of mIP with an overlap, we saved the precise spatial locations of all identified definite microbleeds on 7T scans of the 10 subjects and used them to establish "ground truth" ratings. Each subject was rated 8 times (ie, for 2 conditions of mIP processing by 4 raters). A lesion on a particular location was labeled as a definite microbleed for the ground truth rating if it had been labeled as such on at least 5 of the 8 ratings. Lesion locations that had been marked as definite microbleeds on 4 or fewer of the 8 ratings were examined in a consensus meeting with 3 raters (J.d.B., M.B., G.J.B.) and were included or excluded from the ground truth rating on the basis of the majority of votes.

Statistical Analysis
Intra-and inter-rater agreement of the number of rated microbleeds (possible, definite, and total [possible ϩ definite]) were calculated as an ICC for the 2/0 and 4/Ϫ2 mIP processed scans of the 10 subjects.
The effect of an overlap in the mIP processing of 7T scans was assessed by comparing 2/0 with 4/Ϫ2 mIP processing within the 10 subjects. Individual definite microbleed locations were compared between rated scans of each rater with the ground truth ratings to determine true-positives, false-positives, and false-negatives. These values were used to calculate a sensitivity and SI. The sensitivity and SI were compared between 2/0 and 4/Ϫ2 processing.

Intra-and Inter-Rater Agreement
The intrarater agreement for rating the number of microbleeds on 7T scans was good for definite (ICC, mean of 2/0 and 4/Ϫ2: 0.73) and total microbleeds (0.74), but poor for possible microbleeds (0.33). The inter-rater agreement for rating the number of microbleeds on 7T scans was moderate for definite microbleeds (0.46), lower for total microbleeds (0.38), and poor for possible microbleeds (0.06).

Effect of Using an Overlap in mIP Processing
The ground truth ratings of the 10 subjects included a median number of 3 definite microbleeds (range, 0 -7), with a total of 36 definite microbleeds, of which 31 were lobar, 4 were deep, and 1 was cerebellar. The Table presents the ratings of individual definite microbleed locations per rater compared with the ground truth rating. Mean rating times of all raters between scans with 2/0 and 4/Ϫ2 mIP processing were comparable (29 minutes per scan on average). For individual definite microbleed locations for each rater compared with the locations in the ground truth ratings, 4/Ϫ2 mIP (mean sensitivity/SI of all raters, 0.60/0.67) had a slightly higher sensitivity and SI than 2/0 mIP (0.58/0.63). Microbleed localization (lobar or deep) only had a limited effect on the reliability of visual detection.

DISCUSSION
Improvement in the detection of microbleeds is important to further our understanding of this novel marker of cerebral smallvessel disease. With 7T dual-echo T2*-weighted MR imaging smaller and up to 3 times as many microbleeds can be detected than on 1.5T MR imaging. 2 In contrast to lower resolution scans, at 7T, MR image processing by means of mIP is required to aid the visual rating and reduce the large number of sections. 2 We showed that scans with mIP processing of 2/0 and 4/Ϫ2 showed an optimal balance be-tween the number of detected microbleeds and the number of sections and thus the rating time. Furthermore, detection of microbleeds could benefit from mIP processing with an overlap. Other factors (ie, pulse sequence, sequence parameters, spatial resolution, and the use of susceptibility-weighted imaging) could also influence the reliability of microbleed detection 4 and should be standardized within studies and even between studies to allow comparison of results.
Even under optimal mIP conditions, inter-rater agreement of the visual rating of microbleeds on 7T MR imaging was only moderate and lower than the inter-rater agreement for the same raters at 1.5T. The major source of disagreement between raters was that raters missed microbleeds. This is probably due to the limited size of some of these lesions, susceptibility effects of other structures, 1,4 and the large number of sections for every subject. Disagreement was more marked for possible than for definite microbleeds, and the rating of possible microbleeds on 7T scans proved  to be unreliable. The accuracy of the rating of the 1.5T scans fell well within the range of previous studies (eg, Gregoire et al 3 ), indicating that the disagreement was attributable to the properties of the 7T scans rather than to poor performance of the raters.

CONCLUSIONS
7T MR imaging improves cerebral microbleed detection and will further our understanding of the pathophysiology of these lesions. Visual rating of microbleeds at 7T MR imaging should preferably involve multiple trained raters to ensure optimal lesion detection. At present, microbleed detection at 7T MR imaging should, in our view, be primarily used in a research setting. Further studies are needed to establish the clinical relevance of microbleeds on 7T MR imaging. Moreover, rating times need to be reduced to ensure further dissemination of the technique. This could be achieved by automated or semiautomated image analysis methods, which are currently becoming available. 5