<p>Depression is a major global public health issue that profoundly deteriorates the quality of life for patients and increases the risk of mortality. Electroencephalogram (EEG) signals, as objective physiological markers, have emerged as a focal point in research for depression detection. However, existing methods suffer from inadequate feature representation and fusion, as well as poor model generalization performance due to individual variability. Inspired by the above observations, we propose a cross-subject depression detection method based on the synergy of dynamic adaptive feature fusion and domain adaptation. Specifically, we design a CNN-Transformer dual-branch structure to separately capture local and global EEG features, and further introduce a lightweight dynamically adaptive attention fusion module to efficiently integrate multi-branch features. Moreover, we propose an end-to-end collaborative optimization framework that unifies feature learning and domain adaptation. By jointly optimizing the source domain classification loss, pseudo-label loss, and domain-adversarial loss, the model extracts discriminative representations while aligning cross-domain distributions, thereby achieving deep coupling between feature discriminability and domain invariance and effectively improving cross-subject generalization. Experiments are conducted on two public datasets. Our method achieves 94.39% and 90.77% accuracy, outperforming the state-of-the-art baselines by 5.94% and 2.75%, respectively.</p>