脑梗死是全球最常见的致残和死亡原因之一，在中风症状出现时，病变的定位和体积对临床诊断至关重要。近年来，许多研究在自动病变分割领域提出了表现优异的方法和网络。但脑梗死病变阶段和数据模态的不同，脑梗死区大小、形态差异较大，故目前研究提出的分割网络一般只针对特定阶段和数据模态脑梗死数据，而在实际临床使用时，脑梗死数据所在阶段、模态类型都是未知的，且可能有设备硬件等限制，需要考虑较为通用的模型。
本文选择脑组织分割领域较为经典和通用的模型，包括FCN、U-Net、3D U-Net，并对参数较多的3D U-Net进行优化，加入残差模块得到3D残差U-Net，在相同数据集上训练四种网络进行评估，给出网络选择的建议。网络训练时使用ISLES 2022数据集，对数据集中DWI、ADC两种模态的MRI数据进行缩放、归一化等预处理，在计算每个样本中病变占比，减少无标记的样本输入网络，缓解数据不平衡的问题。用DICE、精确度、召回率、ASSD、HD、体积差评估网络，结果表明FCN、U-Net、3D U-Net、3D残差U-Net的准确率依次上升，说明对3D U-Net的优化有效，在追求分割精度时，优先选择3D残差U-Net，在设备性能有限时，优先选择U-Net。
 

Cerebral infarction is one of the most common causes of disability and death worldwide, and location and volume of lesions are critical to clinical diagnosis at the onset of stroke symptoms. In recent years, many studies have presented excellent methods and networks in the field of automatic lesion segmentation. However, the size and shape of cerebral infarction area vary greatly due to the different stages and data modes of cerebral infarction. Therefore, the segmentation network proposed in the current study generally only targets at specific stages and data modes of cerebral infarction data. However, in the actual clinical use, the stage and mode type of cerebral infarction data are unknown, and there may be equipment and hardware limitations, so a more general model should be considered.
In this paper, classic and universal models in the field of brain tissue segmentation, including FCN, U-Net and 3D U-Net, are selected, and 3D U-Net with more parameters is optimized, and the residual module is added to obtain 3D residual U-Net. Four kinds of networks are trained on the same data set for evaluation, and suggestions for network selection are given. During network training, ISLES 2022 data set is used to pre-process the MRI data of DWI and ADC modes in the data set, such as scaling and normalization, so as to calculate the lesion proportion in each sample, reduce the input of unlabeled samples to the network, and alleviate the problem of data imbalance. DICE, accuracy, recall rate, ASSD, HD and volume difference were used to evaluate the network. The results showed that the accuracy of FCN, U-Net, 3D U-Net and 3D residual U-Net increased successively, indicating that the optimization of 3D U-Net was effective, and 3D residual U-Net was preferred when the segmentation accuracy was pursued. If the device performance is limited, U-Net is preferred.