【摘要】 目的 探究应用基于注意力机制的深度学习网络在头部MR三点非对称回波水脂分离成像（IDEAL）图像上进行5种组织成分（实质、脑脊液、颅骨、空气、软组织）分割的可行性。方法 收集2019年9月至2021年8月杭州市临平区第一人民医院40名健康志愿者的头部CT图像和MR IDEAL图像。训练集包含30个样本，共1 784张图像；测试集包含10个样本，共618张图像。将分割结果与标注区域进行对比评估网络模型效能。利用Elastix软件配准CT图像和MR IDEAL图像；基于CT值得到骨和空气区域的金标准，同时使用SPM12工具包从IDEAL图像得到脑实质、脑脊液区域的金标准，头部其余部分视为软组织区域的金标准；最后，采用最轻量的Segformer模型，将IDEAL的水图、脂图、同相图组成三通道输入到网络模型中进行训练，实现头颅组织成分（实质、脑脊液、颅骨、空气、软组织）的分割。采用Dice相似性系数（DSC）、像素准确度（PA）、均交并比（IoU）评价Segformer在各个组织成分上的分割性能。结果 测试集的各组织成分（空气、骨骼、脑实质、脑脊液、软组织）占比分别为0.125±0.016、0.184±0.015、0.375±0.019、0.085±0.011、0.231±0.020。测试集的DSC为0.822±0.039,PA为0.931±0.015,IoU为0.714±0.050。结论 利用基于注意力机制的深度学习网络能够实现IDEAL图像的头部组织成分分割。更多还原

【Abstract】 Objective To explore the feasibility of tissue segmentation on head MR iterative decomposition of water and fat with echo asymmetric and least squares estimation(IDEAL) images with deep learning network based on attention mechanism. Methods Head CT and MR IDEAL images of 40 healthy subjects September 2019 to August 2021 from the First People’s Hospital of Linping District were analyzed. The training set included 30 samples with 1 784 images and the test set included 10 samples with 618 images. The efficiency of the model was evaluated by comparing the segmentation results with the gold standards. Elastix software was used to register CT and MR IDEAL images, and then the gold standard for bone and air areas was labeled based on CT values. SPM12 was used to obtain the brain parenchyma and cerebrospinal fluid regions from the IDEAL in-phase images, and the rest of the head was considered as the gold standard of soft tissue regions. Finally, the water, fat and in-phase images of IDEAL were composed of three channels and input into the network for training. MIT-B0, the lightest Segformer model, was used to realize the segmentation of head tissue components(brain parenchyma, cerebrospinal fluid, skull, air and soft tissue). Results The weighted average Dice similarity-coefficient of the test set was 0.822±0.039, weighted average pixel accuracy was 0.931±0.015, and weighted average intersection over union was 0.714±0.050. Conclusion Deep learning network based on attention mechanism could realize efficient head tissue components segmentation on MR IDEAL image.更多还原