Individualized lesion-symptom mapping using explainable artificial intelligence for the cognitive impact of white...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Cognitive Neuroscience

Background:

• Cerebral small vessel disease (CSVD) causes lesions impacting cognition.
• Current lesion-symptom mapping (LSM) methods, like support vector regression (SVR), primarily offer group-level insights.
• There is a need for individual-level LSM to understand lesion-cognition relationships.

Purpose of the Study:

• To introduce a novel, individual-level LSM approach using CNNs and XAI.
• To map the relationship between vascular lesions and cognitive impairment on a patient-specific basis.
• To evaluate the performance of the proposed method against existing techniques.

Main Methods:

• A CNN was employed to predict cognitive scores from lesion data.
• Explainable AI (XAI) techniques were integrated with the CNN to generate lesion attribution maps.
• The method was validated using simulated cognitive data and real white matter hyperintensity (WMH) maps from 821 patients, benchmarked against SVR and a fully connected neural network (FNN).

Main Results:

• The CNN demonstrated high predictive performance on simulated data (R² = 0.964).
• CNN with XAI successfully generated patient-specific attribution maps, accurately highlighting critical lesion locations.
• On real patient data, SVR showed higher predictive performance (R² = 0.291) than the CNN (R² = 0.216), though both outperformed total WMH volume (R² = 0.013).

Conclusions:

• CNNs combined with XAI offer a viable method for individual-level lesion-symptom mapping.
• The approach generates patient-specific attribution maps, providing valuable insights into lesion-cognition relationships.
• Further development of this AI-driven LSM technique holds promise for clinical applications.