Efficient detection of cerebral microbleeds on 7.0T MR images using the radial symmetry transform Hugo J. Kuijf a,* , Jeroen de Bresser a,b , Mirjam I. Geerlings c , Mandy M.A. Conijn c,d , Max A. Viergever a , Geert Jan Biessels b , Koen L. Vincken a a Image Sciences Institute, University Medical Center Utrecht, Utrecht, The Netherlands b Department of Neurology, Rudolf Magnus Institute of Neuroscience, University Medical Center Utrecht, Utrecht, The Netherlands c Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, The Netherlands d Department of Radiology, University Medical Center Utrecht, Utrecht, The Netherlands Abstract Cerebral microbleeds (CMBs) are commonly detected on MRI and have recently received an increased interest, because they are associated with vascular disease and dementia. Identification and rating of CMBs on MRI images may be facilitated by semi-automatic detection, particularly on high-resolution images acquired at high field strength. For these images, visual rating is time-consuming and has limited reproducibility. We present the radial symmetry transform (RST) as an efficient method for semi-automated CMB detection on 7.0T MR images, with a high sensitivity and a low number of false positives that have to be censored manually. The RST was computed on both echoes of a dual-echo T2*-weighted gradient echo 7.0T MR sequence in 18 partici- pants from the Second Manifestations of ARTerial disease (SMART) study. Potential CMBs were identified by combining the output of the transform on both echoes. Each potential CMB identified through the RST was visually checked by two raters to identify probable CMBs. The scoring time needed to manually reject false positives was recorded. The sensitivity of 71.2% is higher than that of individual human raters on 7.0T scans and the required human rater time is reduced from 30 to 2 minutes per scan on average. The RST outperforms published semi-automated methods in terms of either a higher sensitivity or less false positives, and requires much less human rater time. Keywords: cerebral microbleeds, radial symmetry transform, 7.0T brain MRI 1. Introduction Interest in cerebral microbleeds (CMBs) is increasing rapidly since a few years. CMBs are seen as phenomena distinct from larger hemorrhages. CMBs are associated with hypertensive vasculopathy, white matter hyperinten- sities and lacunar infarcts, and they are a key MRI marker of cerebral amyloid angiopathy (Knudsen et al. (2001); Wardlaw et al. (2006); Vernooij et al. (2008); Greenberg et al. (2009); Theysohn et al. (2011)). CMBs consist of hemosiderin deposits (Fazekas et al. (1999)) that are para- magnetic and cause a local susceptibility effect inside the magnetic field of the MR scanner. As a result, CMBs can be visualized as round, hypointense spots on a T2*- weighted gradient echo MR sequence. At regular field strength (1–3T), CMBs are usually defined as having a diameter ranging from 2 to 10 mm (Cordonnier et al. (2007)). The current standard for microbleed detection is vi- sual rating with validated visual rating scales (Greenberg * Corresponding author. Heidelberglaan 100, Room Q0S.459, 3584 CX Utrecht, The Netherlands. Email address: hugok@isi.uu.nl (Hugo J. Kuijf) et al. (2009); Gregoire et al. (2009)). As visual rating is time-consuming and has limited reproducibility, (semi- )automated detection may improve rating quality and de- crease rating time. Recently, two methods on semi-automatic detection of CMBs have been published by Seghier et al. (2011) and Barnes et al. (2011). Seghier described a method using a unified segmentation-normalization approach to detect microbleeds. The method identified 77% of patients with microbleeds; no results were given on detection of the actual individual microbleeds. While numbers of false positives were not reported, manual removal of the false positives required 5–10 minutes on average. Barnes used a combination of statistical thresholding and a support vec- tor machine supervised learning classifier on susceptibility weighted images. This method detected 81.7% of all in- dividual microbleeds present in their data (identifying all patients). On average, over 100 false positives were found per patient, which takes a human rater 5–15 minutes to remove. With the introduction of high-field 7.0T MR scanners, detection of much smaller CMBs has become feasible (Conijn et al. (2010); De Reuck et al. (2011)). However, while visual rating may be suitable for scans acquired at reg- ular field strength, it takes a single rater about 30 min- Preprint submitted to NeuroImage September 5, 2013