Multi-output lesion-symptom mapping using deep learning and explainable AI in small vessel disease

Abstract

Cerebral small vessel disease causes cognitive impairment, dementia, and stroke and is characterized by white matter hyperintensities (WMH) and other brain lesions. Lesion-symptom mapping (LSM) aims to understand the relationship between brain lesion location and cognition by identifying strategic lesion locations. This study presents a multi-output deep learning lesion-symptom mapping (DL-LSM) approach using explainable artificial intelligence (XAI). This approach is validated in a simulation study using WMH segmentations of 821 memory clinic patients and artificial cognitive scores. The study comprised three experiments. The first involved generating artificial cognitive scores based on the lesion load within three predefined regions of interest (ROIs). The second experiment studied the impact of adding noise to these scores on the DL and XAI methods. The third experiment explored whether intercorrelations between different ROIs in the artificial cognitive scores could be detected. Two convolutional neural networks (CNN) were developed to predict the artificial cognitive scores, and XAI was used to identify the locations that influenced these predictions. The methods were evaluated by quantifying the model’s predictive performance, identifying the ROIs in the XAI's attribution maps, and quantifying the intercorrelation of the detected ROIs. This study demonstrates that DL models can predict multiple artificial cognitive scores based on WMH segmentations, and that XAI can identify the ROIs associated with the simulated cognitive scores. Additionally, the results demonstrate that DL-LSM is robust to low levels of noise in the artificial cognitive scores and can detect intercorrelations between ROIs. These findings indicate that DL and XAI can be used to perform LSM in order to predict cognitive scores and determine their relationship with specific lesion locations.