基于U-Net的颅内出血识别算法

针对颅脑计算机断层成像（CT）影像中脑出血的分析和识别，提出采用神经网络模型U-Net与轮廓识别相结合的方法提取脑实质区域，通过阈值分割算法分析血块的图像纹理特征，并过滤软组织、脑组织和脑脊液等无关生理组织结构，实现对颅内出血点的精确定位，最后采用插值方法将出血区域进行三维重建，对血块的三维形态作出评估.在天津市某医疗机构提供的500例颅脑CT数据上进行了验证测试，实验结果表明，该算法达到97.4%的目标识别准确率，能够为脑出血诊断提供参考.

U-Net Based Intracranial Hemorrhage Recognition

Aiming at analysis and recognition of cerebral hemorrhage from craniocerebral computed tomography (CT) images, a method combining neural network model U-Net with contour recognition is proposed to extract brain parenchymal regions. The image texture features of hemorrhagic area are firstly extracted by adaptive threshold segmentation algorithm, the precise location of hemorrhagic area is thereafter obtained through filtering the irrelevant physiological tissues such as soft tissue, brain tissue and cerebrospinal fluid. Finally, a three-dimensional structure of the hemorrhagic area is reconstructed based on interpolation to evaluate the amount of hemorrhage. A validation test with 500 patients from medical institution in Tianjin shows that the algorithm achieves a target recognition accuracy of 97.4% and could provide reference for the diagnosis of cerebral hemorrhage.