基于深度学习的脑卒中病灶分割系统研究

Study of the stroke lesion segmentation system based on deep learning

为了改善人工描绘缺血性脑卒中病灶的主观差异性,提升诊断缺血性脑卒中的速率和精确度,本研究基于Attention U-Net深度学习模型搭建深度学习自动分割系统,将DWI、ADC等多模态磁共振影像作为系统输入并提取病灶的多层次特征,获得自动分割结果。结果显示,该系统的Dice可达到0.91,IoU达到0.93,远远优于U-Net算法。该系统可以辅助放射科医生进行脑卒中缺血区域定位,从而改善患者的临床结局。

In order to improve the subjective difference of manual depiction of ischemic stroke lesions and enhance the diagnosis rate and accuracy of ischemic stroke,in this paper,a deep learning automatic segmentation system is built based on the Attention U-Net deep learning model,which takes multimodal magnetic resonance images such as DWI and ADC as the system input and extracts the multilevel features of lesions to obtain automatic segmentation results.The results show that the Dice of the system can reach 0.91,and the IoU can reach 0.93,which is far superior to the U-Net algorithm.The system can assist radiologists in locating ischemic areas of stroke,thus improving the clinical outcomes of patients.