Neurovista: A bidirectional masked cross-Modal fusion network for robust EEG-to-Image decoding

Electroencephalography (EEG)-based visual decoding has significant potential in brain-computer interfaces but faces substantial challenges due to noise, inter-subject variability, and limited fine-grained alignment between neural signals and visual representations. Existing approaches predominantly utilize global EEG embeddings and static fusion methods, restricting their capability to capture nuanced cross-modal interactions. To address these limitations, We propose NeuroVista, a novel framework that integrates localized EEG masking with dynamic bidirectional cross-modal attention, achieving state-of-the-art EEG-to-image decoding performance. Specifically, NeuroVista employs a channel-level EEG masking strategy during training, encouraging the model to learn robust, context-sensitive neural features, thus significantly improving generalization and noise resistance. Simultaneously, our bidirectional cross-modal attention module dynamically aligns EEG embeddings with corresponding visual features, enhancing semantic coherence across modalities. Extensive experiments on standard EEG-to-image benchmarks demonstrate that NeuroVista consistently outperforms state-of-the-art methods, achieving up to +16.0 % top-1 accuracy improvement in both subject-dependent and subject-independent settings. Our results validate the effectiveness of combining localized masking and interactive cross-modal attention, establishing NeuroVista as a robust, interpretable, and highly generalizable approach for EEG-based visual decoding tasks.