中心博士生赵麾宇同学的工作——Deep learning-based automated segmentation for the quantitative diagnosis of cerebral small vessel disease via multisequence MRI（基于深度学习的多序列MRI分割实现定量化多病灶脑小血管病诊断）的相关成果最近被Frontiers in Neurology期刊接收，该工作得到了国家自然科学基金（T2225023和82271337）部分资助。
现有的脑小血管疾病（CSVD）的视觉评分系统无法准确和定量评估全球病变负荷。本研究旨在开发一种基于深度学习（DL）的自动分割方法，以量化多级磁共振成像（MRI）上CSVD的典型神经成像标记。本研究分析了内部MRI扫描（2018年7月至2022年7月）和外部（2012年11月至2015年1月）数据集。开发了一种基于DL的分割方法，以根据分段结果评估白质高强度（WMH），脑微观散布（CMB），腔体空间（EPVS）的定量体积。骰子和其他定量指标用于访问DL分割结果。 Pearson相关系数用于相关分析，并且通过方差分析（ANOVA）评估了不同视觉评分之间标记体积的差异。最后，计算定量Z评分代表与CSVD相关的大脑负担。
本研究评估了总共105例内部患者（64.8±7.4岁，70名男性）和58例外部患者（68.2±6.8岁，29名男性）。内部数据集中WMH，CMB，lac和EPVS的骰子值分别为0.85、0.74、0.76和0.75。 DL与手动方法结果之间的正相关性非常好（总体Pearson相关性= 0.968、0.978、0.948和0.947）。 CSVD神经成像标记的预测体积显示出不同视觉分数的组之间的显着差异（p <0.001）。反映CSVD全球负担的定量Z分数也与公认的总负担评分（p <0.001）息息相关。
摘要:
Objective:Existing visual scoring systems for cerebral small vessel disease (CSVD) cannot assess the global lesion load accurately and quantitatively. We aimed to develop an automated segmentation method based on deep learning (DL) to quantify the typical neuroimaging markers of CSVD on multisequence magnetic resonance imaging (MRI).
Materials and methods:MRI scans from internal (July 2018 to July 2022) and external (November 2012 to January 2015) datasets were analyzed. A DL-based segmentation method was developed to evaluate the quantitative volumes of white matter hyperintensity (WMH), cerebral microbleeds (CMBs), lacunes, and enlarged perivascular spaces (EPVSs) according to the segmentation results. Dice and other quantitative metrics were used to access the DL segmentation results. Pearson correlation coefficients were used for correlation analysis, and the differences in marker volumes among different visual scores were assessed via analysis of variance (ANOVA). Finally, a quantitative Z score was calculated to represent CSVD-related brain burden.
Results:A total of 105 internal patients (64.8 ± 7.4 years, 70 males) and 58 external patients (68.2 ± 6.8 years, 29 males) were evaluated. The Dice values for WMH, CMBs, lacunes, and EPVSs in the internal dataset were 0.85, 0.74, 0.76, and 0.75, respectively. The positive correlation between the DL and the manual approach results was excellent (overall Pearson correlation = 0.968, 0.978, 0.948, and 0.947, respectively). The predicted volumes of the CSVD neuroimaging markers showed significant differences among the groups with different visual scores (p < 0.001). The quantitative Z scores reflecting CSVD global burden also correlated well with the widely recognized total burden score (p < 0.001).
Conclusion: An automated DL model was developed for the segmentation of four CSVD neuroimaging markers on multisequence MRI, providing a strong basis for further CSVD research.