SDCRNet: A Lightweight Spatial Detail Fused Context Reconstruction Network for Real-time Polyp Segmentation

The blurring boundaries and large variations of digestive polyps with complicated anatomical environments severely degrade the performance of automatic polyp segmentation.Though current deep learning-based polyp segmentation methods achieve competitive results, most of these methods often face significant challenges when dealing with out-of-distribution data and real-time performance in clinical applications.In this study, we propose a new lightweight spatial detail fused context reconstruction network (SDCRNet) for real-time polyp segmentation aiming to obtain satisfactory segmenting accuracy while avoiding large-scale computation.Our SDCRNet builds a novel lightweight ConvPFormer block as MetaFormer encoder baseline and then constructs a new multi-resolution spatial detail fused context reconstruction (SDCR) to effectively reconstruct the lost textures and spatial detail features through layer-by-layer interaction of the shallow spatial details with the deep global context, which avoids maintaining high-resolution network propagation while capturing accurate local information and boundary features.Extensive experiments on four benchmark datasets show our SDCRNet obtains satisfactory segmenting results and real-time performance compared with the state-of-the-art methods.