"Standard MRI shows white matter hyperintensities, but these are common in healthy elderly people too," explains Professor Tang. "DTI captures microstructural integrity of white matter – fractional anisotropy (FA) and mean diffusivity (MD) – which is much more sensitive to the subtle damage caused by small vessel disease." The team built a DenseNet – a densely connected convolutional neural network – that learns hierarchical features directly from raw DTI volumes. They trained and validated it on an internal dataset of 134 SVCI and 171 SIVD patients. To ensure the model works on data from different scanners and populations, they used unsupervised domain adaptation (UDA) with an external cohort of 90 SVCI and 103 SIVD patients. Only the unlabeled target‑domain images were used for adaptation; an independent test set (45 SVCI, 50 SIVD) was held out for final evaluation.
The best model, using FA and MD together, achieved an accuracy of 0.902 on the internal test set and 0.926 on the target‑domain test set, with area under the ROC curve (AUC) of 0.951 and 0.942 respectively. "This is substantially better than our previous machine‑learning approach that required manual feature engineering," notes Professor Li. "More importantly, the model's predicted SVCI probability correlates strongly with actual neuropsychological scores – MoCA, MMSE, memory recall and Trail Making tests – so the probability itself gives clinicians a continuous measure of cognitive severity."
To make the "black box" interpretable, the researchers used guided backpropagation to generate salient maps – voxel‑wise contributions to the model's decision. The model focused on 11 white matter regions, including the corona radiata, superior longitudinal fasciculus, corpus callosum, internal capsule and posterior thalamic radiation. These are precisely the tracts known to be damaged in small vessel disease and are functionally linked to attention, executive control and memory. "When we took only the DTI data within those salient regions and trained a separate CNN to predict neuropsychological scores, the predictions were significantly correlated with true scores – proving that the model's attention is neuropsychologically meaningful," says Professor Qin.
SVCI is heterogeneous: one patient may have mainly memory problems, another executive dysfunction. The team wanted to go beyond binary diagnosis to an imaging‑based tool that could estimate an individual's risk in each cognitive domain. They first computed voxel‑wise mutual information (MI) between the FA‑MD images and six neuropsychological scales in SVCI patients. This produced six domain‑specific "relevance maps" showing where white matter damage carries the most information about cognitive performance. Overlap between these maps was low (mean Dice 0.057), confirming that each cognitive domain has a distinct structural footprint.
For each SVCI patient, the researchers then calculated the Structural Similarity Index Measure (SSIM) between the patient's own salient weight map (from the DenseNet) and each of the six domain‑specific MI maps. "SSIM tells us how much the patient's pattern of white matter abnormality resembles the population‑level pattern that is most informative for, say, memory or executive function," explains Dr. He. Using unsupervised K‑means clustering on the SSIM scores, they divided patients into low, moderate and high similarity subgroups for each cognitive domain. In every domain, patients in the high‑similarity cluster had significantly worse neuropsychological performance – for example, lower MoCA scores or longer Trail Making times. "This means we can now stratify SVCI patients not just by overall diagnosis, but by their predicted risk of impairment in specific cognitive functions, all from a single DTI scan."
The framework requires only a standard DTI sequence, which is already part of many clinical MRI protocols. It does not rely on time‑consuming neuropsychological testing, making it particularly valuable for elderly or resource‑limited settings. The authors acknowledge limitations: the cohort size is modest for deep learning, and the current analysis is cross‑sectional. However, their VIVA cohort is undergoing annual follow‑up, which will enable longitudinal prediction of cognitive decline. Future work will incorporate multimodal imaging and blood biomarkers.
"Our study shows that deep learning can not only accurately identify SVCI from SIVD, but also extract a personalized 'cognitive risk signature' from the same brain scan," concludes Professor Tang. "This moves us closer to precision medicine for vascular cognitive impairment – enabling early, targeted interventions tailored to each patient's unique pattern of white matter injury."
Authors of the paper include Miao He, Yunsi Yin, Junda Qu, Yan Wang, Xinwei Que, Xinyi Xia, Tongtong Zhang, Jiangting Li, Junyi Shen, Weihong Song, Qi Qin, Chunlin Li, and Yi Tang.
This work was supported by the Capital's Funds for Health Improvement and Research (2024-2-1032), National Key Research, Development Program of China (2022YFC3602600 and 2023YFF1203502), National Natural Science Foundation of China (82201568 and 62576226), Beijing Natural Science Foundation (L251023), Beijing Nova Program (20240484566), Beijing Outstanding Young Scientist Program (JWZQ20240101023), STI2030-Major Projects (2021ZD0201801), Non-profit Central Research Institute Fund of Chinese Academy of Medical Science (2024-JKCS-12), and Xuanwu Hospital Talent Convergence Program (HZ2025PYYX005).
The paper "Deep Learning for Classifying and Cognitive Profiling of Subcortical Vascular Cognitive Impairment" was published in the journal Cyborg and Bionic Systems on May. 13, 2026, at DOI: 10.34133/cbsystems.0561.
