基于自注意力卷积网络对脑卒中术后血肿预测

Prediction of Hematoma after Stroke Based on Self-Attention Convolutional Network

摘要: 本文针对出血性脑卒中手术后血肿扩张的问题，建立自注意力卷积神经网络模型预测其扩张概率。该模型依据患者的个人史、疾病史、治疗后的影像信息及其相关特征等因素，预测所有患者在手术后24 H内发生血肿体积扩大的概率。自注意力机制可以捕捉输入序列中的长距离依赖关系。相比于传统的卷积神经网络，它不受局部感受野的限制，能够在整个输入序列上进行全局的关联建模，预测精度达到97.5%。该模型在一定程度可以帮助医护人员判断患者术后的恢复情况。

Abstract: This paper aims at the occurrence of hematoma dilatation after hemorrhagic stroke. A self-attention convolutional neural network model is established to predict its expansion probability. Based on the patient’s personal history, disease history, post-treatment imaging information and related features, the model predicted the probability of hematoma enlargement within 24 H after surgery for all patients. Self-attention mechanisms can capture long-distance dependencies in input sequences. Compared with the traditional convolutional neural network, it is not limited by the local receptive field, and can conduct global correlation modeling on the whole input sequence, and the prediction accuracy reaches 97.5%. To some extent, this model can help medical staff to judge the recovery of patients after surgery.