In order to detect the cerebral microbleed (CMB) voxels within brain, we used susceptibility weighted imaging to scan the subjects. Then, we used undersampling to solve the accuracy paradox caused from the imbalanced data between CMB voxels and non-CMB voxels. we developed a seven-layer deep neural network (DNN), which includes one input layer, four sparse autoencoder layers, one softmax layer, and one output layer. Our simulation showed this method achieved a sensitivity of 95.13%, a specificity of 93.33%, and an accuracy of 94.23%. The result is better than three state-of-the-art approaches.

Introduction

Cerebral microbleed (CMB) [49] is small foci of chronic blood products in normal brain tissues. They are closely related with glomerular filtration [35], dementia [55], cortical superficial siderosis [26], and ageing [16]. They are important recognized entity with the rapid development of magnetic resonance imaging (MRI) especially the susceptibility weighted imaging (SWI). The hemosiderin within CMB foci is superparamagnetic, which causes significant local inhomogeneity in the magnetic field around CMB, leading fast decay of MRI signal. Hence, CMB appear hypointensity in the scanned image. Traditional interpretation depends on the MARS (microbleed anatomical rating scale) [22] that draws up stringent rules to classify CMB into two types: Bdefinite^ and Bpossible^ [3]. Nevertheless, the manual interpretation are not reliable due to the high intra-observer and interobserver variability. Visual screening is prone to either confuse with CMB mimics or miss small CMBs [46]. In the last decade, computer scientists tried to solve this problem based on computer vision and image processing techniques. Fazlollahi (2015) [20] combined multi-scale mechanism and Laplacian of Gaussian approach. They abbreviated it as MSLoG. They also used random forest (abbreviated as RF) classifiers. Seghier (2011) [54] proposed a microbleed detection via automated segmentation (MIDAS) technique. Barnes (2011) [4] relied on a statistical thresholding algorithm to detect the hypointensity. They then used support vector machine (SVM) classifier to separate true CMB from others. Bian (2013) [6] employed a 2D fast RST to detect putative CMBs. Afterwards, false results were removed using features of geometry. Kuijf (2012) [27] presented a radial symmetry transform (RST) method. Charidimou (2012) [8] discussed the principles, methodologies, and rational of CMB and its mapping in vascular dementia. Bai (2013) [2] detected CMBs in super-acute ischemic stroke patients treated with intravenous thrombolysis. Roy (2015) [52] proposed a novel multiple radial symmetry transform (MRST) and RF method. Chen (2016) [9] used leaky rectified linear unit (LReLU). Hou (2016) [24] proposed a four-layer deep neural network (DNN) method. Nevertheless, the detection accuracy of above methods are still quite low. For example: Bai’s method [2] combined multi-modality imaging, but they did not use computer vision approach to increase the identification performance. Roy’s method [52] obtained a sensitivity of 85.7%, which is quite higher than human interpretation, but it did not explore the power of computer vision fully.