Learning Subject-Invariant Representation for Brain Decoding via Multi-Region Integration and Adversarial Alignment

Brain decoding aims to translate neural signals into meaningful information, advancing brain-computer interfaces and cognitive neuroscience theories. Existing methods typically focus on signals from the human visual cortex and require retraining or subject-specific parameters for each individual, limiting their generalizability. In this work, we present CortexLink, a unified brain decoding framework designed to learn subject-invariant representations across multiple subjects and brain regions without introducing any subject-specific parameters. Our method integrates four key components: 1) a non-parametric voxel clustering strategy that standardizes signal dimensionality across subjects, 2) multi-region feature integration to fuse information from visual and language-related cortices, 3) adversarial training to extract subject-invariant representations, and 4) dual perceptual-semantic guidance to capture the multifaceted nature of brain responses. Extensive experiments on the Natural Scenes Dataset benchmark demonstrate that CortexLink achieves state-of-the-art performance across three settings. It outperforms existing methods in multi-subject joint training scenarios. Under the challenging leave-one-subject-out (LOSO) setting, CortexLink directly decodes brain activity from held-out subjects without subject-specific adaptation, providing evidence that the proposed parameter-free design can support unseen-subject decoding within the benchmark setting. In addition, auxiliary experiments conducted on the GOD and THINGS-fMRI datasets further verify the portability and robustness of the proposed subject-invariant components across different datasets beyond NSD. Code is available at https://github.com/tqxg2018/CortexLink