<p>Accurate bronchial lumen identification is essential for ensuring the quality and safety of bronchoscopic diagnosis and interventions. Traditional clinical practice relies extensively on manual interpretation by experienced endoscopists. While convolutional neural network (CNN)-based methods have improved efficiency in normal lumen localization, their performance may degrade in the presence of lesion-induced variations, such as structural distortion and inflammatory artifacts. To address these challenges, we propose a frequency-guided deep network with cross-domain attention (FGCANet) for bronchial lumen classification under complex conditions. The model integrates a spatial feature extraction backbone with an auxiliary frequency-guided branch, in which a learnable frequency filter suppresses high-frequency noise while preserving critical structural information. Furthermore, a cross-domain attention mechanism facilitates the adaptive fusion of frequency and spatial features, thereby enhancing discriminative representation learning. Evaluated on a retrospective real-world dataset containing abnormal bronchial lumen images, FGCANet achieves competitive performance compared to several representative CNN- and Transformer-based models, with more consistent performance across both normal and abnormal cases.</p>