Deep learning framework improves diagnosis of vascular cognitive impairment

Subcortical ischemic vascular disease (SIVD) caused by cerebral small vessel disease is commonly characterized by white matter hyperintensities and multiple lacunar infarcts, and a significant proportion of patients gradually develop cognitive decline and progress to subcortical vascular cognitive impairment (SVCI). Early differentiation between SVCI and SIVD without cognitive impairment is important to slow cognitive decline and guide intervention strategies. However, current SVCI diagnosis typically relies on a combined assessment of clinical symptoms, structural MRI, and neuropsychological scales, which can be time-consuming and susceptible to assessment conditions and subjective biases, especially in older adults and in resource-limited settings. Additionally, white matter hyperintensity on conventional structural MRI is also common in older adults, limiting its specificity and sensitivity for early microstructural white matter damage.

Diffusion tensor imaging (DTI) can provide more sensitive information about white matter microstructure, while deep learning has the potential to automatically extract disease-relevant imaging features. Therefore, using DTI and interpretable deep learning to accurately identify SVCI and further profile individual risk across cognitive domains has become an important direction to advance accurate diagnosis and personalized intervention in SVCI. ”

Miao He, author, researcher at Capital Medical University

In this study, we developed a diffusion tensor imaging (DTI)-based deep learning framework to identify subcortical vascular cognitive impairment (SVCI) and further stratify multidomain cognitive risk in patients with subcortical ischemic vascular disease (SIVD). The researchers collected data on DTI scans and neuropsychological scales from an internal cohort of 134 cognitively unimpaired SVCI patients and 171 SIVD patients, and used an external community cohort of 90 SVCI patients and 103 SIVD patients for unsupervised domain adaptation and independent testing. After preprocessing, the DTI images were converted into white matter microstructure metrics including FA, MD, AD, and RD and input into the DenseNet model for SVCI classification. We applied an unsupervised domain adaptation strategy to reduce distribution differences between datasets and improve model generalization for external data. The researchers then used saliency maps to identify key white matter regions contributing to model decisions and computed mutual information maps between DTI images and six neuropsychological scales, including MMSE, MoCA, immediate recall, delayed recall, TMT-A, and TMT-B. Finally, by measuring the structural similarity between each individual’s saliency map and domain-specific mutual information map and applying unsupervised clustering, we stratified SVCI patients into low-risk, intermediate-risk, and high-risk subgroups by cognitive domain.

The results showed that the DTI-based DenseNet model can accurately distinguish between SVCI and SIVD patients without cognitive impairment and maintains good generalization of external data. On the internal test set, the model achieved an accuracy of 0.902. After incorporating unsupervised domain adaptation, its accuracy reached 0.926 on the target domain test set and the AUC was 0.942, showing stable performance across different image sources. The probability of SVCI generated by the model was significantly associated with multiple neuropsychological scales, including MoCA, MMSE, immediate recall, delayed recall, and TMT-A/TMT-B, suggesting that the predictive output is not only useful for classification but also reflects the severity of cognitive impairment. Furthermore, saliency map analysis showed that the model determination mainly relied on white matter tracts such as the corona radiata, corpus callosum, posterior limb of the internal capsule, superior longitudinal fasciculus, retrothalamic radiation, and external capsule, with the corona radiata contributing most prominently. These regions are closely associated with memory, executive function, attention, and visuospatial deficits commonly observed in SVCI. For cognitive profiling, researchers found distinct white matter association patterns across different neuropsychological scales and used structural similarities between individual saliency and mutual information maps to stratify SVCI patients into low-, intermediate-, and high-risk subgroups by cognitive domain. Patients with higher similarity had worse cognitive performance, indicating that this framework can further support individualized multidomain cognitive risk stratification.

The importance of this study lies not only in using DTI and DenseNet to accurately differentiate between SVCI and SIVD, but also in moving imaging AI beyond disease classification to personalized cognitive risk profiling. Through unsupervised domain adaptation, the model maintained good generalization to external data. By combining saliency and mutual information maps, this study further showed that the white matter regions highlighted by the model are neuropsychologically meaningful and may reflect impairment risk across cognitive domains such as memory, executive function, and attention. This approach provides clinicians with a more objective and scalable complementary tool, especially in settings where comprehensive neuropsychological testing is difficult to perform, and provides new directions for accurate stratification and individualized intervention for patients with SVCI. At the same time, some limitations remain. The sample size is still relatively modest for deep learning training, generalizability across centers, scanners, and populations requires large-scale validation, and current analyzes are primarily cross-sectional, so it is not yet possible to directly predict future cognitive decline at the individual level. Additionally, cognitive risk subgroups require further validation with long-term follow-up and multimodal imaging.

“Future studies incorporating larger multicenter longitudinal datasets, functional images, and blood biomarkers may further enhance the clinical value of this framework for accurate diagnosis and intervention guidance in SVCI,” Miao He said.

Authors of this paper include Miao He, Yunsi ying, Junda Qu, Yan Wang, Xinwei Que, Xinyi Xia, Tongtong Zhang, Jianting Li, Junyi Shen, Weihong Song, Qi Qin, Chunlin Li, and Yi Tang.