What Do Persistent Misclassifications Tell Us About Alzheimer's Disease Detection using Structural MRI?

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Neurology

Background:

• Deep learning (DL) models show promise in detecting Alzheimer's Disease (AD) using structural MRI (sMRI).
• Systematic analysis of DL model failure modes in AD detection is limited.
• Understanding misclassification patterns is crucial for improving AI diagnostic tools.

Purpose of the Study:

• To investigate persistent misclassification patterns in DL models for AD detection.
• To characterize subjects who are consistently misclassified by DL models.
• To explore whether persistent false negatives in AD detection are due to disease staging or atypical presentation.

Main Methods:

• Trained two DL architectures to classify AD from cognitively normal (CN) participants using sMRI data from the ADNI dataset.
• Examined persistent misclassifications across 100 model instances and various training configurations.
• Analyzed longitudinal sMRI scans to assess changes in DL model predictions over time.

Main Results:

• Identified a subgroup of subjects persistently misclassified across multiple DL models.
• These persistently misclassified subjects showed distinct atrophy patterns, including hippocampal-sparing and minimal atrophy subtypes.
• Changes in DL model predictions (false negative to true positive) occurred in a subset of subjects over intervals up to five years.

Conclusions:

• Persistent misclassification in DL-based AD detection may not solely be explained by disease progression or staging.
• Disease heterogeneity, characterized by distinct atrophy subtypes, significantly impacts DL model performance.
• Clinical AI model development for AD detection must consider and address disease variability for improved accuracy and reliability.