Retrospective detail reconstruction network for mitigating shallow information loss in colorectal polyp segmentation

Abstract

Abstract Existing medical polyp segmentation networks predominantly rely on hierarchical feature representations to improve boundary delineation. However, we identify a critical yet underexplored issue, termed shallow information loss, wherein low-level layers irreversibly suppress low-contrast but semantically essential edge cues during forward propagation, while attenuated gradients in backpropagation are insufficient to recover such early-stage information loss. This problem is especially severe in polyps exhibiting fractal-like structural characteristics, where cross-scale self-similarity is progressively disrupted across spatial resolutions, ultimately degrading segmentation performance. Moreover, most existing approaches attempt to learn a direct pixel-to-semantic mapping in a single step, lacking progressive feature guidance and resulting in inadequate global semantic awareness. To address these limitations, we propose RDNet, a retrospective detail network composed of a backward detail reconstruction (BDR) module and a cascaded synergistic optimization (CSO) module. The BDR module injects discriminative semantic information backward in a layer-wise manner to recover weak-response regions that are suppressed during forward propagation, while cross-layer feature calibration enables accurate reactivation of shallow structural details. The CSO module further decouples boundary and region representations and employs a conditional gating mechanism to selectively activate complementary features, thereby enhancing structural consistency and semantic discrimination. Extensive experiments conducted on five public benchmark datasets demonstrate that RDNet consistently outperforms state-of-the-art methods, with particularly significant improvements in fractal-like multi-polyp segmentation scenarios, validating its robustness and effectiveness in complex clinical environments.