Colorectal cancer (CRC) screening critically relies on precise polyp segmentation. However, existing methodologies often struggle with lesion heterogeneity, indistinct boundaries, and elevated rates of missed diagnoses. In response to these challenges, this study proposes BiGAMR-Net, a novel deep learning framework that synergistically combines Bidirectional Gated Attention with Multi-scale Residual Learning to enhance segmentation performance. The proposed architecture comprises three core components: (1) a channel-wise Squeeze-and-Excitation attention mechanism designed to suppress irrelevant noise and emphasize discriminative features; (2) Multi-scale Residual (MR) blocks, which utilize parallel atrous convolutions to capture diverse receptive fields while maintaining stable gradient propagation; and (3) A Bidirectional Fusion Module (BFM) facilitates dynamic, cross-scale feature integration to align spatial and semantic information. Comprehensive evaluations across five benchmark datasets (NBI, Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, ETIS) demonstrate that BiGAMR-Net achieves state-of-the-art performance, with mean Dice scores of 0.956 on Kvasir-SEG and 0.958 on ETIS exceeding the performance of competitive models such as nnUNet and DuckNet by up to 4.4%. The framework performs robustly in challenging scenarios, including narrow-band imaging and complex mucosal patterns. Ablation studies further validate the effectiveness of the BFM in refining lesion boundaries (3.8% Dice improvement) and the MR module in detecting small lesions (2.9% IoU gain). These findings underscore the clinical promise of BiGAMR-Net in reducing diagnostic oversights and improving the reliability of CRC screening. Future research will focus on enabling real-time inference, enhancing cross-modality generalization, and conducting large-scale clinical trials to support broader clinical translation.