Source-Free Active Learning for Adapting Alzheimer's Diagnostic Deep Learning Models Across Neuroimaging Cohorts.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Alzheimer's Disease (AD) classification across multiple neuroimaging sites is hindered by domain shifts due to data heterogeneity.
• Developing robust diagnostic models requires addressing variations in data acquisition, devices, and demographics.

Purpose of the Study:

• To propose an uncertainty-informed active learning framework for Source-Free (SF) Domain Adaptation (DA) in AD classification.
• To improve model generalization across diverse neuroimaging studies without requiring source domain data during deployment.

Main Methods:

• Leveraged Monte Carlo dropout for estimating prediction uncertainty to guide sample selection from the target domain.
• Employed an active learning strategy to adapt models using informative target samples.
• Evaluated the framework on a large-scale dataset (3,177 participants) from five neuroimaging studies using regional brain volume measurements.

Main Results:

• Achieved the highest median Area Under the Curve (AUC) of 91.4% across all source-target combinations, outperforming baseline models (89.7%).
• Demonstrated superior performance compared to other SF and Source-Aware (SA) DA methods.
• Quantified distribution shifts using maximum mean discrepancy to assess effectiveness under varying inter-site shifts.

Conclusions:

• The proposed uncertainty-based active learning framework effectively addresses domain shifts in multi-site neuroimaging for AD classification.
• SF methods can achieve performance comparable or superior to SA approaches while respecting data privacy constraints.
• This approach enhances the robustness and generalizability of AD diagnostic models across different neuroimaging studies.