Wavelet-guided Multi-scale Semantic Fusion for Gastrointestinal Lesion Detection

Deep learning has shown promise in image pattern recognition and is expected to enhance diagnostic efficiency and accuracy in gastrointestinal (GI) endoscopy, particularly for wireless capsule endoscopy (WCE), which produces 60,000–80,000 images per examination. However, robust lesion detection remains challenging due to the inherent characteristics of GI images, such as repetitive textures, low lesion-to-background contrast, limited image resolution, and the scarcity of large-scale annotated clinical datasets. In this paper, multi-scale features are extracted from GI images and fused in a context-guided manner to optimize the network’s ability to learn lesion visual patterns. Specifically, an improved Haar Wavelet Downsampling (HWD) module is proposed to better preserve lesion features in low-resolution images, while a Selective Boundary Aggregation (SBA) mechanism is introduced to enhance multi-scale semantic fusion. These modules are implemented and integrated into a modified backbone similar to YOLOv12. The proposed network is evaluated on the public LDPolypVideo dataset. Experimental results suggest that it outperforms competing models, achieving improvements of 6.5%, 4.4%, and 5.0% in recall, mAP@50, and F1-score, respectively, while maintaining a favorable balance between detection performance and computational efficiency.