Explainable artificial intelligence for neuroimaging-based dementia diagnosis and prognosis

Area of Science:

• Neuroimaging and Artificial Intelligence
• Machine Learning in Medicine

Background:

• Accurate dementia prediction using AI and neuroimaging is possible, but 'black box' models lack trust.
• Explainable AI (XAI) offers insights into model behavior and feature importance.
• Vision Transformers (ViT) present a self-explainable alternative to Convolutional Neural Networks (CNNs).

Purpose of the Study:

• To compare the effectiveness of ten XAI methods across CNN and ViT architectures.
• To assess the utility of XAI in understanding dementia prediction models.
• To evaluate XAI for Alzheimer's disease (AD) diagnosis and mild cognitive impairment (MCI) to AD conversion prognosis.

Main Methods:

• T1-weighted MRI data were used to train classification and prognosis models.
• Ten XAI techniques were systematically evaluated.
• Model performance was assessed for AD diagnosis and MCI-to-AD conversion prediction.

Main Results:

• Models achieved 81% balanced accuracy for AD diagnosis and 67% for MCI prognosis.
• XAI outputs successfully identified brain regions critical for AD.
• XAI provided valuable insights for predicting MCI to AD conversion.

Conclusions:

• XAI methods can validate that AI models utilize relevant neuroimaging features.
• XAI generates valuable data for further clinical analysis and trust in AI.
• Explainable AI enhances the interpretability and reliability of AI-driven neuroimaging analysis for dementia.