Cerebral Hemorrhage Recognition Based on Mask R-CNN Network

AbstractThis paper presents a method to identify cerebral hemorrhage on CT images aiming at providing support for diagnosis. Firstly, we use Mask R-CNN network to segment the brain parenchyma area. Secondly, we use threshold segmentation algorithm to locate the blood clot area. At last, a 3D visualization method of cerebral hemorrhage is built by cross-sectional contours interpolation. In terms of detection accuracy, 294 brain CT image sets are sampled, which included 113 of healthy brains and 181 of intracranial hematoma. We divided the above 294 samples into two categories, one with cerebral hemorrhage symptoms, and the other with no cerebral hemorrhage symptoms. We use the accuracy of binary classification to evaluate the disease diagnosis effect of the algorithm, which is 97.6% by statistics. At the same time, in order to evaluate the effect of threshold segmentation in cerebral hemorrhage, we use the relative error index. 1000 slices were randomly selected from 294 brain CT data. The maximum relative error and average relative error of pixel threshold segmentation of cerebral hemorrhage points were calculated on 1000 slices, which are 5% and 0.84%. Based on the above experimental results, we believe that this method can provide a reference for the diagnosis of cerebral hemorrhage.
Source: Sensing and Imaging - Category: Biomedical Engineering Source Type: research