深度学习超分辨重建常规厚层MRI用于海马定量影像诊断的可行性分析

Feasibility analysis of quantitative hippocampal diagnosis with deep-learning based super-resolution reconstruction of clinical thick-slice MRI

目的:评价深度学习超分辨重建常规厚层MRI应用于海马形态学定量分析的可行性。方法:纳入东部战区总医院同时具有厚层常规临床脑T1WI及三维各向同性高分辨图像的404例被试的影像学资料。其中,230例MRI阴性癫痫数据用于训练集,121例颞叶癫痫伴海马硬化(TLE-HS)患者及53名健康人(NC)数据作为测试集,行基于深度学习SCSR的超分辨重建处理。以各向同性高分辨率图像数据作为真实对照,采用基于体素和基于皮层的形态学分析方法,分别计算测试集重建图像和真实图像的灰质体积、海马体积及皮层厚度指标,采用皮尔逊相关分析比较正常对照组重建图像与真实图像间海马体积的相关性,并采用Dice系数评价重建图像和真实图像间组间比较显著差异区域的重合度。结果:正常对照组重建图像与真实图像海马体积的相关性分别为0.90(左侧,P<0.001)和0.91(右侧,P<0.001)。重建图像与真实图像间颞叶癫痫患者组和正常对照组灰质体积的差异区域的Dice系数分别为0.74(左侧TLE-HS vs.正常对照)和0.74(右侧TLE-HS vs.正常对照),皮层厚度的差异区域的Dice系数分别是0.50(左侧TLE-HS vs.正常对照)和0.59(右侧TLE-HS vs.正常对照)。结论:本研究证明了基于深度学习的超分辨算法辅助临床诊断数据应用于疾病定量化分析的可靠性。

Objective:To evaluate the reliability of clinical thick-slice MRI applied to quantitative analysis of hippocampal morphology with deep learning based super-resolution.Methods:A total of 404 subjects with both clinical thick-slice brain T1 WI and threedimensional isotropic high-resolution images at the Jinling hospital were recruited.230 MRI negative epilepsy patients were set as training set,121 temporal lobe epilepsy patients with hippocampal sclerosis(TLE-HS)and 53 normal controls(NC)as the test set,deep learning based superresolution reconstruction by using the structure constrained super resolution network(SCSR).By setting isotropic high-resolution images as ground truth,the gray matter volume,hippocampal volume,and cortical thickness of the reconstructed images and ground truth from the test set were evaluated using voxel-and surface-based morphometry,respectively.Pearson correlation analysis was used to compare hippocampal volume between the reconstructed images and ground truth from normal controls,and the Dice coefficient was used to evaluate the similarity between reconstructed images and ground truth in group comparisons.Results:Correlations between reconstructed and real image hippocampal volumes in normal controls were 0.90(left,P<0.001)and 0.91(right,P<0.001).The dice coefficients of difference in gray matter volume between the reconstructed images and ground truth for the TLE-HS patients and NC were 0.74(left TLE-HS vs.NC)and 0.74(right TLE-HS vs.NC),respectively,and the difference in cortical thickness were 0.50(left TLE-HS vs.NC)and 0.59(right TLE-HS vs.NC).Conclusion:Our study demonstrates the reliability of a deep learning based super-resolution algorithm to assist clinical diagnostic data for quantitative analysis of diseases.