Multi-output Deep Learning and Explainable AI methods for Lesion-Symptom Mapping

Abstract

Cerebral small vessel disease is a common condition that can lead to stroke and dementia, which cause disability in different cognitive domains. Lesion-Symptom Mapping (LSM) techniques aim to find correlations between the affected locations of the brain and cognition decline. Voxel-based and Support Vector Regression LSM are the current golden standards, but they have limitations in terms of model complexity and interpretability. Deep Learning (DL) methods have the potential of building complex models, which can be useful to tackle this challenge. This study proposes a DL pipeline that can predict multiple artificial cognition scores from lesion MR images and correctly identify the brain locations that are most relevant to make specific score predictions using explainable Artificial Intelligence (xAI) methods. The study analyzes the performance of single and multioutput models and explores different xAI methods (Integrated Gradients, Gradient Shap and Occlusion) to understand the information each can provide. Overall, this project demonstrates that DL techniques can satisfactorily predict multiple regression outputs from segmented lesion MR images and identify the key regions that affect each score, which could be used in the field of LSM to understand the underlying brain mechanisms that contribute to various neurological and psychiatric disorders.