Résumé: Automatic polyp segmentation can improve the accuracy of colonoscopy and plays a crucial role in colorectal cancer prevention. However, existing U-shaped convolutional neural networks fail to satisfactorily localize the boundaries for polyp region, which inevitably degenerates the performance of polyp segmentation. In this article, we propose a boundary learning and enhancement network (BLE-Net) that finely restores edge localization by combining two novel boundary modules. Specifically, a novel boundary learning (BL) module is deployed on the encoder stage to embed edge details into high-level features via a bottom-up fusion way, thereby producing discriminative features with both semantics and boundary information. Moreover, to strengthen the weak responses at fuzzy boundaries, we further design a boundary enhancement (BE) module, in which three cascaded boundary-aware attention blocks progressively endow ambiguous edge cues and rectify preceding maps in a coarse-to-fine fashion. Extensive experimental results on five polyp datasets demonstrate that BLE-Net has excellent segmentation performance and generalization capability, outperforming the state-of-the-arts.

Langue: Anglais