Scholarlink

<p>The mechanism linking spatial heterogeneity of perivascular spaces (PVS) to cognitive impairment has remained unclear. This study aims to develop an interpretable deep learning method to predict cognitive impairment in patients with cerebral small vessel disease (CSVD) and to elucidate the association between brain region-specific PVS distribution and model decision patterns. This retrospective study enrolled 279 patients with arteriosclerotic CSVD from Huashan Hospital, Fudan University, collecting their T1-weighted MRI images and cognitive assessment data. The Montreal cognitive assessment (MoCA) scores were used to dichotomize cognitive status 26 for normal cognition [NC], &lt; 26 for cognitive impairment [CI]) for binary prediction of cognitive states. A dual-stream deep learning model was constructed to simultaneously process raw MRI and PVS-segmented images. Multimodal features were integrated through a modified ResNet architecture with embedded gradient-weighted class activation mapping (Grad-CAM) for decision visualization. PVS volume and heatmap intensity were quantified across brain regions. Dice value was used as an evaluation metric for the segmentation method. Accuracy, precision, specificity, sensitivity, and F1-score were used as classification evaluation criteria to assess the classification performance. Spatial correlations were evaluated using Spearman’s rank analysis. The Dice value of the segmentation method is 0.729. The classification method achieved accuracies of 0.847 and 0.804 on internal and external test sets, respectively, presenting improvements of 8.2 and 9.0 percentage points over single-modality inputs. The precision, specificity, sensitivity, and F1-score in the internal test set were 0.787, 0.821, 0.881, and 0.831, respectively. Grad-CAM visualization revealed 0.841 spatial overlap between PVS and high-intensity heatmap regions. The five brain regions with the largest PVS volumes were basal ganglia (415.14 ± 28.67) mm<sup>3</sup>, periventricular white matter (219.37 ± 19.24) mm<sup>3</sup>, frontal lobe (72.54 ± 6.53) mm<sup>3</sup>, insula (58.56 ± 5.12) mm<sup>3</sup>, and thalamus (45.45 ± 4.01) mm<sup>3</sup>. The five regions with the largest weights of the thermogram were the frontal lobe (26.45% ± 2.31%), periventricular white matter (17.08% ± 1.52%), parietal lobe (15.83% ± 1.42%), insula (9.36% ± 0.84%), and basal ganglia (8.88% ± 0.79%). Significant spatial correlations were observed at both individual patient level (mean <i>r</i> = 0.628, <i>p</i> &lt; 0.05) and population level (global <i>r</i> = 0.645, <i>p</i> = 0.451). The impact of PVS spatial heterogeneity on cognitive impairment exhibits significant brain region specificity. Our explainable deep learning framework provides a novel tool for early screening and mechanistic investigation of CSVD-related cognitive decline.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning Reveals the Association Between Perivascular Space Distribution and Cognitive Impairment in Cerebral Small Vessel Disease

  • 期刊论文

  • 发表日期：

    2025年12月30日

查看全文
  • Luyuan Jin,
  • Miaoyi Zhang,
  • Zhihong Zou,
  • Jie Tang,
  • Huiyu Zhao,
  • Jianhui Fu,
  • Weiwen Zou

摘要

The mechanism linking spatial heterogeneity of perivascular spaces (PVS) to cognitive impairment has remained unclear. This study aims to develop an interpretable deep learning method to predict cognitive impairment in patients with cerebral small vessel disease (CSVD) and to elucidate the association between brain region-specific PVS distribution and model decision patterns. This retrospective study enrolled 279 patients with arteriosclerotic CSVD from Huashan Hospital, Fudan University, collecting their T1-weighted MRI images and cognitive assessment data. The Montreal cognitive assessment (MoCA) scores were used to dichotomize cognitive status 26 for normal cognition [NC], < 26 for cognitive impairment [CI]) for binary prediction of cognitive states. A dual-stream deep learning model was constructed to simultaneously process raw MRI and PVS-segmented images. Multimodal features were integrated through a modified ResNet architecture with embedded gradient-weighted class activation mapping (Grad-CAM) for decision visualization. PVS volume and heatmap intensity were quantified across brain regions. Dice value was used as an evaluation metric for the segmentation method. Accuracy, precision, specificity, sensitivity, and F1-score were used as classification evaluation criteria to assess the classification performance. Spatial correlations were evaluated using Spearman’s rank analysis. The Dice value of the segmentation method is 0.729. The classification method achieved accuracies of 0.847 and 0.804 on internal and external test sets, respectively, presenting improvements of 8.2 and 9.0 percentage points over single-modality inputs. The precision, specificity, sensitivity, and F1-score in the internal test set were 0.787, 0.821, 0.881, and 0.831, respectively. Grad-CAM visualization revealed 0.841 spatial overlap between PVS and high-intensity heatmap regions. The five brain regions with the largest PVS volumes were basal ganglia (415.14 ± 28.67) mm3, periventricular white matter (219.37 ± 19.24) mm3, frontal lobe (72.54 ± 6.53) mm3, insula (58.56 ± 5.12) mm3, and thalamus (45.45 ± 4.01) mm3. The five regions with the largest weights of the thermogram were the frontal lobe (26.45% ± 2.31%), periventricular white matter (17.08% ± 1.52%), parietal lobe (15.83% ± 1.42%), insula (9.36% ± 0.84%), and basal ganglia (8.88% ± 0.79%). Significant spatial correlations were observed at both individual patient level (mean r = 0.628, p < 0.05) and population level (global r = 0.645, p = 0.451). The impact of PVS spatial heterogeneity on cognitive impairment exhibits significant brain region specificity. Our explainable deep learning framework provides a novel tool for early screening and mechanistic investigation of CSVD-related cognitive decline.