This paper presents a novel machine learning approach for computer-aided detection of microbleeds in SWI. The major contributions are: identifying microbleed extent in order to extract proper cubic regions-of-interest (ROI) containing the structure, (2) extracting a set of robust 3- dimensional (3D) Radon- and Hessian-based shape descriptors within the ROIs as well as 2D Radon features computed on intensity-projection images of the corresponding ROIs, and (3) incorporating a cascade of random forests (RF) classifiers to iteratively reduce false detection rates while maintaining a high sensitivity.