SIG-CFFNet: Structural Information-Guided Cascaded Feature Fusion Network for Gastrointestinal Anatomy Classification

Accurate identification of gastrointestinal endoscopic anatomical structures is critical for improving diagnostic accuracy and reducing missed detection rates. However, endoscopic image quality may be compromised by various factors including lesion interference and inadequate bowel preparation, while the morphological similarity of certain anatomical structures further complicates recognition in low-quality images. To address these challenges, we propose a Structural Information-Guided Cascaded Feature Fusion Network (SIG-CFFNet). Our approach leverages anatomical prior knowledge to guide the cascaded fusion of CNN and Transformer branch features, while incorporating Depthwise Over-parameterized Convolutional Layer (DO-Conv) to enhance feature representation and computational efficiency during the fusion process. Comprehensive experimental results demonstrate the superior performance of our method across multiple evaluation scenarios: it achieves classification accuracy of 73.47% and 87.05% for normal and pathological endoscopic anatomical structures, respectively; attains 99.61% and 87.83% accuracy on the Kvasir-Capsule and HyperKvasir public datasets; and maintains robust performance with 84.46% and 80.21% accuracy in cross-domain evaluations (COVID19-CT and ISIC2018). Notably, our model demonstrates highly competitive or near state-of-the-art recall rates across multiple test scenarios, confirming its clinical applicability and robustness for real-world implementation.