Abstract
Accurate identification of late-life depression (LLD) based on brain MRI is crucial for monitoring its clinical progression over time. Existing learning-based LLD studies often suffer from limited (e.g., tens) data, leading to unreliable model training. While using auxiliary data can enlarge the sample size, the inherent heterogeneity between auxiliary and target MRIs with different acquisition settings often hinders generalizability. To this end, we propose a collaborative domain adaptation (CDA) framework for LLD detection with T1-weighted MRIs, leveraging knowledge learned from large-scale auxiliary source domains to a small-sized target domain. The CDA contains a Vision Transformer (ViT) for capturing global MRI representation and a Convolutional Neural Network (CNN) for extracting local features, both pre-trained on 9,544 MRIs from a public cohort. Its training consists of three components: (1) supervised training on labeled source data, with ViT- and CNN-based encoders for feature extraction and two classifiers for prediction; (2) fine-tuning with feature alignment by minimizing the discrepancy between classifier outputs from two branches to make the categorical boundary clearer; and (3) collaborative training on unlabeled target MRIs, leveraging augmented MRI samples to ensure feature consistency. Extensive experiments on T1-weighted MRIs from a total of 238 subjects suggest that CDA outperforms several state-of-the-art approaches, achieving superior classification accuracy and improved cross-domain generalization.