Learning from heterogeneous structural MRI via collaborative domain adaptation for late-Life depression assessment.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Accurate identification of late-life depression (LLD) is vital for patient care.
• Current AI models struggle with limited data and domain differences in brain MRI.
• Cross-domain transferability is a major challenge in applying AI to diverse medical datasets.

Purpose of the Study:

• To develop a robust framework for LLD detection using T1-weighted MRIs.
• To overcome limitations of small sample sizes and domain heterogeneity in LLD identification.
• To improve the generalization of AI models for LLD detection across different clinical sites.

Main Methods:

• Proposed a Collaborative Domain Adaptation (CDA) framework with a dual-branch architecture (Vision Transformer and CNN).
• Employed a three-stage approach: supervised source training, self-supervised target adaptation, and collaborative unlabeled target training.
• Utilized a reliability-aware JSD-based dual-constraint mechanism for pseudo-label filtering and robust prediction consistency.

Main Results:

• The CDA framework demonstrated superior performance compared to state-of-the-art unsupervised domain adaptation methods.
• Experiments on multi-site benchmarks validated the model's enhanced generalization ability.
• The integrated ViT and CNN approach effectively captured both global and local brain MRI features.

Conclusions:

• The CDA framework offers a promising solution for accurate LLD detection despite data limitations and domain shifts.
• This approach facilitates more reliable monitoring and timely intervention for late-life depression.
• CDA shows significant potential for real-world clinical applications in neuroimaging analysis.