QCFSNet: A Quaternion-Guided Inter-Channel and Frequency-Spatial Feature Interaction Strategy for Endoscopic Image Segmentation

In the field of medical consumer electronics, particularly in cervical cancer diagnostics, endoscopic imaging is often affected by uneven illumination and the high similarity in textures and features between lesion regions and surrounding tissues, which severely hinders accurate boundary perception and lesion segmentation. To address these challenges, we propose QCFSNet, a Quaternion-Guided Channel-Frequency-Spatial Feature Interaction framework for segmentation in endoscopic images. This approach incorporates a frequency representation mechanism based on quaternion convolution and wavelet transform to extract high-dimensional and highly discriminative features while alleviating the impact of illumination inconsistency on feature modeling. In addition, a boundary-adaptive attention module is introduced to suppress redundant responses and enhance the contrast between lesion and background regions, while a frequency-guided spatial-channel fusion mechanism strengthens semantic integration across multiple domains. Experiments on a private dataset (CervicalEndoDB) and public benchmarks (CVC-300 and UW-Sinus-Surgery-C/L) demonstrate that the proposed method holds great promise in improving segmentation perception in electronic endoscopic imaging, thereby better supporting downstream clinical tasks such as physician intervention planning.