ABSTRACT
Cerebral microbleeds (CMBs) are increasingly being recognized as an important biomarker for neurovascular diseases. So far, all attempts to count and quantify them have relied on manual methods that are time-consuming and can be inconsistent. A technique is presented that semiautomatically identifies CMBs in susceptibility weighted images (SWI). This will both reduce the processing time and increase the consistency over manual methods. This technique relies on a statistical thresholding algorithm to identify hypointensities within the image. A support vector machine (SVM) supervised learning classifier is then used to separate true CMB from other marked hypointensities. The classifier relies on identifying features such as shape and signal intensity to identify true CMBs. The results from the automated section are then subject to manual review to remove false-positives. This technique is able to achieve a sensitivity of 81.7% compared with the gold standard of manual review and consensus by multiple reviewers. In subjects with many CMBs, this presents a faster alternative to current manual techniques at the cost of some lost sensitivity.
