Recently, emotion recognition utilizing EEG signals as a physiological modality has garnered increasing attention in the research community. EEG offers advantages such as being non-intrusive, inexpensive, portable, and having high temporal resolution, with features across multiple dimensions, including time, frequency, and spatial domains. However, existing studies typically perform cross-domain feature fusion after independently extracting features from different dimensions. n this paper, we introduce the Multidimensional Feature Collaborative Extraction Network (MFCENet). The MFCENet framework primarily comprises a Time-Frequency Feature Extraction Module (TFEM) and a Spatial Feature Extraction Module, with two parallel information streams operating concurrently. First, the spatial feature extraction module learns the channel representations containing spatial features by utilizing the spatial topology and dependency relationships between channels. Then, the time-frequency feature extraction module learns temporal dynamic features while considering spatial relationships through cross-attention. Finally, the extracted multidimensional features are fused and applied for EEG-besed emotion recognition. The proposed approach was assessed using two publicly accessible EEG datasets—DEAP and SEED. The evaluation results validate its effectiveness in accurately recognizing emotional states from EEG signals.