动脉粥样硬化是一种在动脉血管壁上发展起来的复杂的病理过程，一方面可以加重管腔狭窄，另一方面某些易损斑块容易破裂形成血栓导致下游血管栓塞，进而引发脑卒中。多对比度磁共振管壁成像技术联合多个序列进行成像，可以准确识别血管管腔-管壁交界，提供斑块成分信息。目前基于多对比度磁共振影像的颈动脉管壁分割方法主要基于手动或者半自动方法，耗时长，可重复性差。本文以传统U-Net神经网络为基础，分别探讨模型结构、损失函数、数据增强和磁共振影像输入模态对颈动脉管壁分割结果的影响，并结合四个实验中的最优方法得到基于压缩-激励-残差模块（Squeeze-Excitation-Residual Block，SE Res）的U-Net神经网络的多对比度磁共振影像颈动脉管壁分割方法。 在模型结构实验中，基于SE Res U-Net (concatenate)神经网络的分割结果最好，SE Res模块提取特征能力更强，跨层融合中“concatenate”操作的特征表达能力更好；在损失函数实验中，使用Dice损失函数+交叉熵损失函数的分割结果最好，到达最小损失函数值的时间最短，训练过程更稳定；在数据增强实验中，同时使用随机放大后裁剪、随机旋转、随机左右翻转、弹性形变四种数据增强方法的分割结果最好，合理的数据增强方法可以避免模型过拟合，提升分割准确度；在磁共振影像输入模态实验中，使用T1W、T2W、TOF、和MPRAGE四个对比度影像分开卷积然后将特征“concatenate”的输入方法的分割结果最好，该输入方法可以提供更多关于颈动脉管壁的信息，同时保证特征组合更加灵活。 最后，综合以上四个实验的最优方法得到基于SE Res U-Net神经网络的多对比度磁共振影像颈动脉管壁分割方法。该方法在测试集上的颈动脉管壁分割结果显示，灵敏度为0.892，特异性为0.987，Dice系数为0.875，并且和手工分割结果相比，在最大管壁厚度和管壁面积上ICC [95%置信区间]分别为0.937 [0.931-0.942]和0.956 [0.952-0.960]，具有较高一致性。由各种分割示例可以发现，基于SE Res U-Net神经网络模型在各种临床情况下都取得了较好的分割结果。
Atherosclerosis is a complex pathological process that develops on the arterial wall, which can further aggravate lumen stenosis and some vulnerable plaques are prone to rupture and form thrombosis, leading to the embolism of downstream vascular and thereby causing stroke. Multicontrast magnetic resonance vessel wall imaging technology combines with multiple sequences, which can accurately identify the lumen-wall interface and provide composition information of plaque. The segmentation methods of carotid vessel wall based on multicontrast magnetic resonance (MR) images are mainly manual or semi-automatic methods, which are time-consuming and have poor reproducibility. In this paper, we respectively investigated the influence of the model architecture, loss function, data augmentation and input modality of MR images to the segmentation of carotid vessel wall in multicontrast MR images based on the traditional U-Net model. Finally, the model based on the SE Res U-Net neural network was utilized in the segmentation of carotid vessel wall in multicontrast MR images, combining with the optimal methods in the four experiments.In the model architecture experiment, the model based on SE Res U-Net (concatenate) neural network produced the best segmentation results. The ability of feature extraction for SE Res block and feature expression for “concatenate” operation in cross-layer fusion were better. In the loss function experiment, the model using the joint loss function of Dice loss and cross entropy loss produced the best segmentation results, which had the shortest time to reach the minimum loss function and a more stable training process. In the data augmentation experiment, the model using four data augmentation methods simultaneously produced the best segmentation results, which meant reasonable data augmentation methods could avoid overfitting and improve segmentation accuracy. In the input modality of MR images experiment, the model that performed convolution separately for each contrast MR image(T1W, T2W, TOF and MPRAGE) and then concatenated the feature maps produced the best segmentation results. This input modality method can provide more information about carotid vessel wall and combine feature maps more flexibly.Finally, the model based on the SE Res U-Net neural network was utilized in the segmentation of carotid vessel wall in multicontrast MR images, which combining the optimal methods from the above four experiments. The sensitivity, specificity and Dice coefficient of the SE Res U-Net model are 0.892, 0.987, and 0.875 in the test dataset, respectively. The ICC was 0.937[0.931-0.942] for max wall thickness and 0.956[0.952-0.960] for wall area. The segmentation results show substantial agreement for SE Res U-Net segmentation method and manual segmentation method. In addition, SE Res U-Net neural network achieves good segmentation results in various clinical situations.
