Improving Late-Life Depression Analysis with Collaborative Domain Adaptation: Learning from Heterogeneous Structural...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Geriatric Psychiatry

Background:

• Accurate identification of late-life depression (LLD) via brain MRI is vital for clinical monitoring.
• Limited data in existing LLD studies compromises model reliability.
• Heterogeneity in MRI data acquisition across studies hinders generalizability.

Purpose of the Study:

• To propose a collaborative domain adaptation (CDA) framework for LLD detection using T1-weighted MRIs.
• To leverage knowledge from large-scale auxiliary datasets to improve LLD detection in small target datasets.
• To enhance the generalizability of LLD detection models across different MRI acquisition settings.

Main Methods:

• Developed a CDA framework integrating a Vision Transformer (ViT) for global features and a Convolutional Neural Network (CNN) for local features.
• Pre-trained ViT and CNN encoders on 9,544 MRIs from a public cohort.
• Employed supervised training, feature alignment fine-tuning, and collaborative training on unlabeled target MRIs with augmented samples.

Main Results:

• The CDA framework demonstrated superior performance compared to state-of-the-art approaches in LLD detection.
• Achieved higher classification accuracy on T1-weighted MRIs from 238 subjects.
• Showcased improved cross-domain generalization capabilities.

Conclusions:

• The proposed CDA framework effectively addresses data limitations and heterogeneity in LLD MRI studies.
• CDA offers a robust solution for reliable and generalizable LLD detection.
• This approach holds promise for advancing the clinical monitoring of late-life depression.