As one of the most critical visual tasks in autonomous driving, lane detection often suffers from inaccurate lane localization due to obstruction, lighting, and night. To address this issue, current studies propose specialized architectures to enhance lane feature representation. While these methods demonstrate effectiveness, they also increase computational complexity. Therefore, achieving a balance between performance and efficiency remains challenging. Recently, Mamba has demonstrated remarkable capabilities in long-sequence modeling and efficiency. In response, this paper proposes an efficient Mamba-based global Feature Selective Aggregator named MaSA for lane detection. Leveraging Mamba’s selective scanning mechanism and Vision Mamba’s multi-directional scanning strategy, MaSA enables comprehensive information propagation by scanning feature maps in four distinct directions. This approach significantly enhances global context modeling and improves the detection of visually ambiguous lanes. Furthermore, MaSA can be easily integrated with any existing encoder-decoder-based lane detection networks while maintaining a good balance between performance and efficiency. Extensive experiments on TuSimple and CULane datasets demonstrate the effectiveness and compatibility of the proposed method.