Automatic detection of cerebral micro-bleed in SWI images based on 3D CNN

Abstract:
Susceptibility-weighted Imaging (SWI) is an important imaging method in the diagnosis of cerebrovascular diseases in recent years. This technique is of great significance in the diagnosis of Cerbral micro-bleeds (CMBs) in human brain. As a small lesion area, CMBs consumes a lot of time and energy in clinical diagnosis, and the diagnosis results are easily misjudged by the influence of doctors' experience and working conditions. In order to improve the diagnostic efficiency of clinicians, a method for the automatic location and recognition of CMBs in SWI sequence is proposed. Different from the previous methods of CMBs recognition based on low-dimensional features of 2D images or threedimensional shape features extracted manually, this method uses 3D-CNN network to extract three-dimensional features of CMBs leading to a more accurate and efficient detection results. In view of the problem that CMBs are sparse in SWI and difficult to locate, we put forward an improved multiscale Laplace of Gaussian filter (Multi-LOG) algorithm to identify potential CMBs candidate regions. The performance of our method was assessed on 2742 CMBs samples acquired from 480 patients with cerebrovascular diseases at different hospitals. The proposed CMBs detection method achieves an accuracy and a sensitivity of 97.5% and 95.3%, respectively. These performances are relatively high; which demonstrates the robustness and efficiency of our method for automatically detecting and classifying micro-hemorrhage. The proposed deep learning based method may help reduce the clinicians workload in the CMBs detection processes; which will contribute to the precise diagnosis and management of brain diseases.