<p>Finding a safe and feasible path in complex narrow environments has always been a major challenge for mobile robots. The bidirectional rapidly exploring random tree (BiRRT) algorithm, as a variant of the rapidly exploring random tree (RRT) algorithm, significantly improves solving speed but still faces the problem of slow or even infeasible searches in complex narrow environments. To address this issue, we propose an improved BiRRT algorithm combined with deep learning, called neural-guided BiRRT (NG BiRRT). NG BiRRT introduces an intermediate node selection method based on convolutional neural networks (CNN), overcoming the limitations of linear formulas and making the selection of intermediate nodes more efficient and accurate. Based on the intermediate nodes, we propose an adaptive sampling region related to the intermediate node and the start and goal positions to improve the sampling efficiency of the algorithm. Additionally, we adopt a bidirectional search mode with the intermediate node as the root node of the RRT tree, enhancing the efficiency of the algorithm in complex narrow environments. For tree node expansion, we employ a greedy hybrid node growth strategy to increase the probability of generating effective tree nodes in narrow spaces. Meanwhile, safety distance checks are performed to ensure path safety. Simulation results demonstrate that the proposed algorithm can adapt to various complex environments and rapidly return a safe path. Furthermore, real-world experiments further validate the effectiveness of the proposed method.</p>