<p>Accurate segmentation of pneumonia lesions in CT scans is crucial for early diagnosis and effective treatment, particularly during outbreaks of infectious diseases. Existing methods often struggle to detect small lesions and to bridge the semantic gap between encoder and decoder features. In this study, we propose an enhanced hybrid CNN-BiFormer network with residual Haar wavelet downsampling and shared attention skip connection for pneumonia lesion segmentation. The hybrid CNN–BiFormer architecture is designed to model local details and long-range dependencies jointly. To enhance edge feature representation while maintaining computational efficiency, we incorporate a residual Haar wavelet downsampling module, which effectively preserves structural details essential for identifying small lesions. Additionally, a shared attention mechanism is introduced in the skip connections to dynamically align encoder and decoder features, thereby reducing the semantic gap and enhancing segmentation accuracy in challenging regions. We evaluate CBHA-Net on three datasets: COVID-19-CT-Seg, Segmentation dataset nr.2 and a private clinical dataset. The experimental results show that the proposed method achieves competitive performance compared to state-of-the-art models, including TransUNet, U-Net++, and Inf-Net, achieving Dice scores of 84.36%, 82.87%, and 74.68%, respectively. These results suggest that the proposed method exhibits strong segmentation performance and demonstrate its feasibility for use in real clinical settings to aid radiologists in early and accurate pneumonia assessment.</p>