Multioutput lesion-symptom mapping with explainable artificial intelligence in cerebral small vessel disease

Cerebral small vessel disease is a major contributor to future stroke, dementia, and cognitive decline. Lesion-symptom mapping (LSM) studies have established the relationship between the location of brain damage (e.g., white matter hyperintensities) and cognitive outcomes. However, current technology is limited to assessing a single cognitive outcome at a time, whilst cognition is multifaceted with co-occurring and interrelated outcomes. We propose a deep learning-based LSM approach that can do both: predict multiple cognitive outcomes simultaneously and use eXplainable AI (XAI) to generate attribution maps that highlight which brain lesions are involved in these outcomes. We compared two 3D deep learning approaches (a CNN and a residual CNN) and two XAI approaches (occlusion and gradient-SHAP) in a simulation study. Real WMH lesions were obtained from 821 patients, three ROIs were placed in the brain to create impactful locations, and artificial c o gnitive scores were generated based on the lesion presence within these ROIs. Results demonstrate that the deep learning approaches were able to predict the artificial s cores (R ² > 0 .85) a nd r econstruct t he R O Is. U p t o 2 0% n oise can be added to the artificial s cores b efore t he r esults s tart t o d e teriorate. F inally, a rtificial sc ores we re cr eated that depended on multiple ROIs simultaneously, and the CNN method could correctly reconstruct these relationships. Quantitative and qualitative analysis of the XAI-generated attribution maps confirm t hat t he R OI l ocations are correctly highlighted. This proof-of-concept study suggests that CNNs are a feasible approach to multi-outcome LSM.