基于域对齐的深度学习方法在急性缺血性卒中的CT/MRI影像分割上的应用

Application of domain-alignment deep learning methods in CT/MRI image segmentation for acute ischemic stroke

目的基于平扫CT提出一种域对齐方法来显著提高急性缺血性卒中(AIS)的早期快速诊断能力。方法回顾性分析南方医科大学第三附属医院神经内科和神经外科2020年1月~2022年12月收治的入院后3 d内同时接受平扫头颅CT和MRI/DWI、ADC以及T2-Flair序列扫描的AIS患者,构建了一个由318例AIS病例组成的成对CT/MRI影像数据集,分别对每一组配对的教师-学生影像特征进行归一化;再以8∶2的比例随机分为训练集和验证集。设计一种新的生成性对抗性网络来对齐特征层上的跨模式输入,将细节丰富的MRI图像中的语义知识传递到CT图像中进行AIS分割,开发了一种新的域适应算法(Our DA)。结果与目前性能表现较优异的医学影像分割模型nnUNet相比,Our DA明显优于nnU-Net,每一层验证集之间的分割精度提升约15%。结论本研究构建的Our DA模型基于MRI/DWI序列的影像特征并迁移到平扫头颅CT上,对平扫头颅CT上的AIS病灶具有较高的自动分割性能,有助于早期自动识别AIS病灶。

Objective To propose a domain alignment method based on plain scan CT to improve the early and quick diagnosis of acute ischemic stroke(AIS).Methods AIS patients admitted to the Department of Neurology and Neurosurgery of Southern Medical University from January 2020 to December 2022 who received non-contrast head CT,MRI/DWI,ADC and T2-Flair sequence scanning within 3 d after admission were retrospectively analyzed,and a paired CT/MRI image dataset consisting of 318 AIS patients was constructed.The teacher-student image features of each pair were normalized respectively.Then,the training set and the verification set were randomly divided in a ratio of 8∶2.A new generative adversarial network was designed to align cross-mode inputs on the feature layer and transfer semantic knowledge from detailed MRI to CT images for AIS segmentation.A new domain alignment(DA)algorithm(Our DA)was developed.Results Compared with nnUNet,the most powerful medical image segmentation model at present,Our DA was significantly better than nnU-Net,and the segmentation accuracy between each layer verification set was improved by about 15%.Conclusion Our DA model constructed in this study is based on the image features of MRI/DWI sequences and migrated to non-contrast head CT,which has high automatic segmentation performance for AIS lesions on non-contrast head CT,and conducive to early automatic identification of AIS lesions.