Deep Learning for Classifying and Cognitive Profiling of Subcortical Vascular Cognitive Impairment

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Subcortical vascular cognitive impairment (SVCI) presents diagnostic challenges, especially when neuropsychological assessments are difficult.
• Small vessel disease is the primary cause of SVCI, a heterogeneous condition.
• Accurate identification and cognitive profiling are crucial for managing SVCI.

Purpose of the Study:

• To develop a diffusion tensor imaging (DTI)-based DenseNet model for identifying SVCI from subcortical ischemic vascular disease (SIVD).
• To profile multidomain cognitive risks in SVCI patients using DTI-derived information.
• To offer a complementary diagnostic and intervention framework for SVCI.

Main Methods:

• Collected DTI and neuropsychological data from SVCI and SIVD patients for model development and validation.
• Employed unsupervised domain adaptation (UDA) for robust performance on external datasets.
• Utilized DenseNet to generate salient maps, computed mutual information (MI) maps, and structural similarity index measure (SSIM) for cognitive profiling.

Main Results:

• The DTI-based DenseNet achieved high accuracy (0.902 internal, 0.926 target-domain) and AUCs (0.951 and 0.942).
• Salient maps identified key white matter regions associated with SVCI and correlated with neuropsychological performance.
• Cognitive profiling stratified patients into risk subgroups based on domain-specific cognitive impairment.

Conclusions:

• DTI-based DenseNet provides accurate SVCI identification, outperforming traditional methods when assessments are impractical.
• The model enables individualized multi-domain cognitive profiling, revealing structural correlates of cognitive deficits.
• This approach supports clinical diagnosis and personalized treatment strategies for SVCI.