A new study announced on June 9 developed a deep learning framework based on diffusion tensor imaging (DTI) to improve the diagnosis of subcortical vascular cognitive impairment (SVCI) among patients with subcortical ischemic vascular disease (SIVD). The research aimed to address limitations in current diagnostic methods, which often rely on clinical symptoms, structural MRI, and neuropsychological scales that can be time-consuming and susceptible to subjective bias.
Researchers collected DTI scans and neuropsychological data from an internal cohort of 134 SVCI patients and 171 SIVD patients without cognitive impairment. An external community cohort of 90 SVCI patients and 103 SIVD patients was used for unsupervised domain adaptation and independent testing. After preprocessing, DTI images were converted into white matter microstructural metrics—FA, MD, AD, and RD—and input into a DenseNet model for classification. Unsupervised domain adaptation was employed to reduce distribution differences between datasets. Salient maps identified key white matter regions influencing the model's decisions, while mutual information maps linked DTI features with six neuropsychological scales: MMSE, MoCA, Immediate Recall, Delayed Recall, TMT-A, and TMT-B.
The results showed that the DenseNet model could accurately distinguish between SVCI and non-impaired SIVD patients. In internal testing, it achieved an accuracy of 0.902; after applying unsupervised domain adaptation its accuracy reached 0.926 in the target-domain test set with an AUC of 0.942. Model-generated probabilities were significantly associated with multiple neuropsychological scales including MoCA and MMSE scores as well as recall tests and trail-making tasks.
Salient map analysis indicated that decisions mainly relied on white matter tracts such as the corona radiata—which contributed most prominently—the corpus callosum, posterior limb of the internal capsule, superior longitudinal fasciculus, posterior thalamic radiation, and external capsule. These regions are closely related to memory deficits as well as executive function impairments commonly seen in SVCI.
The study also stratified SVCI patients into low-, moderate-, or high-risk subgroups for each cognitive domain by measuring structural similarity between individual salient maps and mutual information maps derived from neuropsychological scale performance. Patients showing higher similarity had worse cognitive outcomes.
"Future studies incorporating larger multicenter longitudinal datasets, functional imaging, and blood biomarkers may further enhance the clinical value of this framework for precision diagnosis and intervention guidance in SVCI," said Miao He.
