基于迁移学习算法对新生儿大脑3D T1WI的灰白质分割及其发育量化研究

Quantitative evaluation of the cortical development on neonates based on segmentation of 3D T1WI images using transfer learning

目的基于机器学习实现新生儿脑T1WI图像分割,并量化评估早产与足月新生儿脑皮层结构的发育状态.材料与方法纳入西安交通大学第一附属医院MRI未见异常的早产新生儿、校正至足月期的早产儿、足月新生儿共计50例.基于外部数据库中的相似数据集,对密集卷积神经网络图像分割模型进行初步训练.然后随机从本地数据中抽取25例作为训练集对模型进行二次训练.同时,基于验证集(样本量:10)的分割效果调整模型参数.最终在15例测试集数据中采用Dice系数对模型分割效果进行评价.基于皮层重建提取双侧大脑半球皮层表面积、厚度以及容积等指标.运用Spearman偏相关分析各指标与校正胎龄的相关性;利用Mann-Whitney U检验进行组间差异分析.结果图像分割模型可有效地实现新生儿脑灰质、白质、脑脊液的分割,Dice系数范围为0.93~0.99.双侧大脑半球皮层表面积、容积与校正胎龄呈正相关(P<0.05);皮层厚度随校正胎龄的变化在左右大脑半球间存在不对称性.除了皮层厚度未见组间差异外,早产新生儿双侧大脑半球皮层表面积、容积均低于足月新生儿(P<0.001).结论通过对密集卷积神经网络训练,可有效地实现新生儿脑T1WI图像分割;基于图像分割与皮层重建可量化评估新生儿脑皮层的发育水平,早产儿双侧大脑半球皮层发育落后于足月儿.

Objective: To implement a machine learning-based segmentation method on neonatal T1WI images and assess the cortical structural maturation of preterm and term neonates. Materials and Methods: This work enrolled 50 subjects without any abnormalities on magnetic resonance imaging from the First Affiliated Hospital of Xi'an Jiaotong University, including preterm neonates, preterm newborns at the term equivalent age, and term neonates. The preliminary training of a densely convolution network was performed by using the similar neonatal dataset from the shared database. Segmentation of the local dataset was performed by using this preliminary model. The segmentation results were modified manually by experts. Then the segmentation model was trained again by using the local data from randomly selected 25 subjects. According to the segmentation quality of the validation dataset (sample size: 10), the model parameters were adjusted. Finally, this segmentation model was assessed by using Dice ratios on the testing dataset (sample size: 15). The surface area, cortical thickness, and cortical volume of left and right hemispheres of the brain were extracted based on the cortical reconstruction. Correlations between these metrics and the postmenstrual age were performed by using the Spearman partial correlation. Inter-group differences were evaluated by using the Mann-Whitney U test. Results: The proposed model could effectively segment the gray matter, white matter, and cerebrospinal fluid regions in T1WI images on neonatal brains. This method was feasible on preterm neonates, preterm newborns at the term equivalent age, and term neonates. The dice ratios ranged from 0.93 to 0.99. Significant positive correlations between surface area, cortical volume and the postmenstrual age were observed on both hemispheres (P<0.05). Postmenstrual age-related change patterns of the cortical thickness on left and right hemispheres were different. Except that no significant inter-group difference could be found in cortical thickness, preterm neonates held smaller surface area and cortical volume on both hemispheres than term neonates (P<0.001). Conclusions: It is feasible to implement segmentation of T1WI images on neonatal brains based on a densely connected convolution network. Metrics extracted from the image segmentation and cortical reconstruction could be used to quantitatively assess the cortical structural maturation of neonates. Cortical maturation was delayed in preterm neonates than term neonates on both hemispheres.