Automatic Detection and Classification of Cerebral Microbleeds Using 3D CNN

Cerebral Microbleeds (CMBs) are referred to tiny foci of hemorrhage in brain parenchyma which are smaller than 5 (to 10) mm in size. The presence of CMBs is implicated in pathophysiology of cognitive impairment, dementia, radiation-induced vascular injury, traumatic brain injury, hypertensive microangiopathy, and aging. On brain Magnetic Resonance Imaging (MRI) scans, CMBs appear as hypointense foci, most notable on T2*-weighted or Susceptibility-Weighted Imaging (SWI). Detecting these tiny microbleeds with naked eye is a difficult and time-consuming task for radiologists. In this study we developed an algorithm for automatic detection of CMBs. We applied a two-step strategy: at first, we applied pre-processed 2D image dataset to You Only Look Once (YOLO V2) for detection of CMBs. Then, these detected CMBs locations are used to segment 3D patches from their original SWI volume in the datasets. Next, these patches are used as inputs for Convolution Neural Network (CNN). In the second step, we reduced the number of False Positives (FP) and improved our classification accuracy using 3D CNN. We used two datasets consisting of 979 patients: 879 of whom for training of models, and the remainder for independent validation. We were able to achieve an accuracy of 81% and reduce the <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>F</mml:mi> <mml:msub><mml:mrow><mml:mi>P</mml:mi></mml:mrow> <mml:mrow><mml:mi>avg</mml:mi></mml:mrow> </mml:msub> </mml:math> to 0.16.